You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
rocksdb/java/rocksjni/comparatorjnicallback.cc

636 lines
21 KiB

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
// This file implements the callback "bridge" between Java and C++ for
// rocksdb::Comparator.
#include "rocksjni/comparatorjnicallback.h"
#include "rocksjni/portal.h"
namespace rocksdb {
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
ComparatorJniCallback::ComparatorJniCallback(
JNIEnv* env, jobject jcomparator,
const ComparatorJniCallbackOptions* options)
: JniCallback(env, jcomparator),
m_options(options) {
// cache the AbstractComparatorJniBridge class as we will reuse it many times for each callback
m_abstract_comparator_jni_bridge_clazz =
static_cast<jclass>(env->NewGlobalRef(AbstractComparatorJniBridge::getJClass(env)));
// Note: The name of a Comparator will not change during it's lifetime,
// so we cache it in a global var
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
jmethodID jname_mid = AbstractComparatorJni::getNameMethodId(env);
if (jname_mid == nullptr) {
// exception thrown: NoSuchMethodException or OutOfMemoryError
return;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
jstring js_name = (jstring)env->CallObjectMethod(m_jcallback_obj, jname_mid);
if (env->ExceptionCheck()) {
// exception thrown
return;
}
jboolean has_exception = JNI_FALSE;
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
m_name = JniUtil::copyString(env, js_name,
&has_exception); // also releases jsName
if (has_exception == JNI_TRUE) {
// exception thrown
return;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
// cache the ByteBuffer class as we will reuse it many times for each callback
m_jbytebuffer_clazz =
static_cast<jclass>(env->NewGlobalRef(ByteBufferJni::getJClass(env)));
m_jcompare_mid = AbstractComparatorJniBridge::getCompareInternalMethodId(
env, m_abstract_comparator_jni_bridge_clazz);
if (m_jcompare_mid == nullptr) {
// exception thrown: NoSuchMethodException or OutOfMemoryError
return;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
m_jshortest_mid =
AbstractComparatorJniBridge::getFindShortestSeparatorInternalMethodId(
env, m_abstract_comparator_jni_bridge_clazz);
if (m_jshortest_mid == nullptr) {
// exception thrown: NoSuchMethodException or OutOfMemoryError
return;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
m_jshort_mid =
AbstractComparatorJniBridge::getFindShortSuccessorInternalMethodId(env,
m_abstract_comparator_jni_bridge_clazz);
if (m_jshort_mid == nullptr) {
// exception thrown: NoSuchMethodException or OutOfMemoryError
return;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
// do we need reusable buffers?
if (m_options->max_reused_buffer_size > -1) {
if (m_options->reused_synchronisation_type
== ReusedSynchronisationType::THREAD_LOCAL) {
// buffers reused per thread
UnrefHandler unref = [](void* ptr) {
ThreadLocalBuf* tlb = reinterpret_cast<ThreadLocalBuf*>(ptr);
jboolean attached_thread = JNI_FALSE;
JNIEnv* _env = JniUtil::getJniEnv(tlb->jvm, &attached_thread);
if (_env != nullptr) {
if (tlb->direct_buffer) {
void* buf = _env->GetDirectBufferAddress(tlb->jbuf);
delete[] static_cast<char*>(buf);
}
_env->DeleteGlobalRef(tlb->jbuf);
JniUtil::releaseJniEnv(tlb->jvm, attached_thread);
}
};
m_tl_buf_a = new ThreadLocalPtr(unref);
m_tl_buf_b = new ThreadLocalPtr(unref);
m_jcompare_buf_a = nullptr;
m_jcompare_buf_b = nullptr;
m_jshortest_buf_start = nullptr;
m_jshortest_buf_limit = nullptr;
m_jshort_buf_key = nullptr;
} else {
//buffers reused and shared across threads
const bool adaptive =
m_options->reused_synchronisation_type == ReusedSynchronisationType::ADAPTIVE_MUTEX;
mtx_compare = std::unique_ptr<port::Mutex>(new port::Mutex(adaptive));
mtx_shortest = std::unique_ptr<port::Mutex>(new port::Mutex(adaptive));
mtx_short = std::unique_ptr<port::Mutex>(new port::Mutex(adaptive));
m_jcompare_buf_a = env->NewGlobalRef(ByteBufferJni::construct(
env, m_options->direct_buffer, m_options->max_reused_buffer_size,
m_jbytebuffer_clazz));
if (m_jcompare_buf_a == nullptr) {
// exception thrown: OutOfMemoryError
return;
}
m_jcompare_buf_b = env->NewGlobalRef(ByteBufferJni::construct(
env, m_options->direct_buffer, m_options->max_reused_buffer_size,
m_jbytebuffer_clazz));
if (m_jcompare_buf_b == nullptr) {
// exception thrown: OutOfMemoryError
return;
}
m_jshortest_buf_start = env->NewGlobalRef(ByteBufferJni::construct(
env, m_options->direct_buffer, m_options->max_reused_buffer_size,
m_jbytebuffer_clazz));
if (m_jshortest_buf_start == nullptr) {
// exception thrown: OutOfMemoryError
return;
}
m_jshortest_buf_limit = env->NewGlobalRef(ByteBufferJni::construct(
env, m_options->direct_buffer, m_options->max_reused_buffer_size,
m_jbytebuffer_clazz));
if (m_jshortest_buf_limit == nullptr) {
// exception thrown: OutOfMemoryError
return;
}
m_jshort_buf_key = env->NewGlobalRef(ByteBufferJni::construct(
env, m_options->direct_buffer, m_options->max_reused_buffer_size,
m_jbytebuffer_clazz));
if (m_jshort_buf_key == nullptr) {
// exception thrown: OutOfMemoryError
return;
}
m_tl_buf_a = nullptr;
m_tl_buf_b = nullptr;
}
} else {
m_jcompare_buf_a = nullptr;
m_jcompare_buf_b = nullptr;
m_jshortest_buf_start = nullptr;
m_jshortest_buf_limit = nullptr;
m_jshort_buf_key = nullptr;
m_tl_buf_a = nullptr;
m_tl_buf_b = nullptr;
}
}
ComparatorJniCallback::~ComparatorJniCallback() {
jboolean attached_thread = JNI_FALSE;
JNIEnv* env = getJniEnv(&attached_thread);
assert(env != nullptr);
env->DeleteGlobalRef(m_abstract_comparator_jni_bridge_clazz);
env->DeleteGlobalRef(m_jbytebuffer_clazz);
if (m_jcompare_buf_a != nullptr) {
if (m_options->direct_buffer) {
void* buf = env->GetDirectBufferAddress(m_jcompare_buf_a);
delete[] static_cast<char*>(buf);
}
env->DeleteGlobalRef(m_jcompare_buf_a);
}
if (m_jcompare_buf_b != nullptr) {
if (m_options->direct_buffer) {
void* buf = env->GetDirectBufferAddress(m_jcompare_buf_b);
delete[] static_cast<char*>(buf);
}
env->DeleteGlobalRef(m_jcompare_buf_b);
}
if (m_jshortest_buf_start != nullptr) {
if (m_options->direct_buffer) {
void* buf = env->GetDirectBufferAddress(m_jshortest_buf_start);
delete[] static_cast<char*>(buf);
}
env->DeleteGlobalRef(m_jshortest_buf_start);
}
if (m_jshortest_buf_limit != nullptr) {
if (m_options->direct_buffer) {
void* buf = env->GetDirectBufferAddress(m_jshortest_buf_limit);
delete[] static_cast<char*>(buf);
}
env->DeleteGlobalRef(m_jshortest_buf_limit);
}
if (m_jshort_buf_key != nullptr) {
if (m_options->direct_buffer) {
void* buf = env->GetDirectBufferAddress(m_jshort_buf_key);
delete[] static_cast<char*>(buf);
}
env->DeleteGlobalRef(m_jshort_buf_key);
}
if (m_tl_buf_a != nullptr) {
delete m_tl_buf_a;
}
if (m_tl_buf_b != nullptr) {
delete m_tl_buf_b;
}
releaseJniEnv(attached_thread);
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
const char* ComparatorJniCallback::Name() const {
return m_name.get();
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
int ComparatorJniCallback::Compare(const Slice& a, const Slice& b) const {
jboolean attached_thread = JNI_FALSE;
JNIEnv* env = getJniEnv(&attached_thread);
assert(env != nullptr);
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
const bool reuse_jbuf_a =
static_cast<int64_t>(a.size()) <= m_options->max_reused_buffer_size;
const bool reuse_jbuf_b =
static_cast<int64_t>(b.size()) <= m_options->max_reused_buffer_size;
MaybeLockForReuse(mtx_compare, reuse_jbuf_a || reuse_jbuf_b);
jobject jcompare_buf_a = GetBuffer(env, a, reuse_jbuf_a, m_tl_buf_a, m_jcompare_buf_a);
if (jcompare_buf_a == nullptr) {
// exception occurred
MaybeUnlockForReuse(mtx_compare, reuse_jbuf_a || reuse_jbuf_b);
env->ExceptionDescribe(); // print out exception to stderr
releaseJniEnv(attached_thread);
return 0;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
jobject jcompare_buf_b = GetBuffer(env, b, reuse_jbuf_b, m_tl_buf_b, m_jcompare_buf_b);
if (jcompare_buf_b == nullptr) {
// exception occurred
if (!reuse_jbuf_a) {
DeleteBuffer(env, jcompare_buf_a);
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
MaybeUnlockForReuse(mtx_compare, reuse_jbuf_a || reuse_jbuf_b);
env->ExceptionDescribe(); // print out exception to stderr
releaseJniEnv(attached_thread);
return 0;
}
jint result =
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
env->CallStaticIntMethod(
m_abstract_comparator_jni_bridge_clazz, m_jcompare_mid,
m_jcallback_obj,
jcompare_buf_a, reuse_jbuf_a ? a.size() : -1,
jcompare_buf_b, reuse_jbuf_b ? b.size() : -1);
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
if (env->ExceptionCheck()) {
// exception thrown from CallIntMethod
env->ExceptionDescribe(); // print out exception to stderr
result = 0; // we could not get a result from java callback so use 0
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
if (!reuse_jbuf_a) {
DeleteBuffer(env, jcompare_buf_a);
}
if (!reuse_jbuf_b) {
DeleteBuffer(env, jcompare_buf_b);
}
MaybeUnlockForReuse(mtx_compare, reuse_jbuf_a || reuse_jbuf_b);
releaseJniEnv(attached_thread);
return result;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
void ComparatorJniCallback::FindShortestSeparator(
std::string* start, const Slice& limit) const {
if (start == nullptr) {
return;
}
jboolean attached_thread = JNI_FALSE;
JNIEnv* env = getJniEnv(&attached_thread);
assert(env != nullptr);
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
const bool reuse_jbuf_start =
static_cast<int64_t>(start->length()) <= m_options->max_reused_buffer_size;
const bool reuse_jbuf_limit =
static_cast<int64_t>(limit.size()) <= m_options->max_reused_buffer_size;
MaybeLockForReuse(mtx_shortest, reuse_jbuf_start || reuse_jbuf_limit);
Slice sstart(start->data(), start->length());
jobject j_start_buf = GetBuffer(env, sstart, reuse_jbuf_start, m_tl_buf_a, m_jshortest_buf_start);
if (j_start_buf == nullptr) {
// exception occurred
MaybeUnlockForReuse(mtx_shortest, reuse_jbuf_start || reuse_jbuf_limit);
env->ExceptionDescribe(); // print out exception to stderr
releaseJniEnv(attached_thread);
return;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
jobject j_limit_buf = GetBuffer(env, limit, reuse_jbuf_limit, m_tl_buf_b, m_jshortest_buf_limit);
if (j_limit_buf == nullptr) {
// exception occurred
if (!reuse_jbuf_start) {
DeleteBuffer(env, j_start_buf);
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
MaybeUnlockForReuse(mtx_shortest, reuse_jbuf_start || reuse_jbuf_limit);
env->ExceptionDescribe(); // print out exception to stderr
releaseJniEnv(attached_thread);
return;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
jint jstart_len = env->CallStaticIntMethod(
m_abstract_comparator_jni_bridge_clazz, m_jshortest_mid,
m_jcallback_obj,
j_start_buf, reuse_jbuf_start ? start->length() : -1,
j_limit_buf, reuse_jbuf_limit ? limit.size() : -1);
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
if (env->ExceptionCheck()) {
// exception thrown from CallIntMethod
env->ExceptionDescribe(); // print out exception to stderr
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
} else if (static_cast<size_t>(jstart_len) != start->length()) {
// start buffer has changed in Java, so update `start` with the result
bool copy_from_non_direct = false;
if (reuse_jbuf_start) {
// reused a buffer
if (m_options->direct_buffer) {
// reused direct buffer
void* start_buf = env->GetDirectBufferAddress(j_start_buf);
if (start_buf == nullptr) {
if (!reuse_jbuf_start) {
DeleteBuffer(env, j_start_buf);
}
if (!reuse_jbuf_limit) {
DeleteBuffer(env, j_limit_buf);
}
MaybeUnlockForReuse(mtx_shortest, reuse_jbuf_start || reuse_jbuf_limit);
rocksdb::RocksDBExceptionJni::ThrowNew(env, "Unable to get Direct Buffer Address");
env->ExceptionDescribe(); // print out exception to stderr
releaseJniEnv(attached_thread);
return;
}
start->assign(static_cast<const char*>(start_buf), jstart_len);
} else {
// reused non-direct buffer
copy_from_non_direct = true;
}
} else {
// there was a new buffer
if (m_options->direct_buffer) {
// it was direct... don't forget to potentially truncate the `start` string
start->resize(jstart_len);
} else {
// it was non-direct
copy_from_non_direct = true;
}
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
if (copy_from_non_direct) {
jbyteArray jarray = ByteBufferJni::array(env, j_start_buf,
m_jbytebuffer_clazz);
if (jarray == nullptr) {
if (!reuse_jbuf_start) {
DeleteBuffer(env, j_start_buf);
}
if (!reuse_jbuf_limit) {
DeleteBuffer(env, j_limit_buf);
}
MaybeUnlockForReuse(mtx_shortest, reuse_jbuf_start || reuse_jbuf_limit);
env->ExceptionDescribe(); // print out exception to stderr
releaseJniEnv(attached_thread);
return;
}
jboolean has_exception = JNI_FALSE;
JniUtil::byteString<std::string>(env, jarray, [start, jstart_len](const char* data, const size_t) {
return start->assign(data, static_cast<size_t>(jstart_len));
}, &has_exception);
env->DeleteLocalRef(jarray);
if (has_exception == JNI_TRUE) {
if (!reuse_jbuf_start) {
DeleteBuffer(env, j_start_buf);
}
if (!reuse_jbuf_limit) {
DeleteBuffer(env, j_limit_buf);
}
env->ExceptionDescribe(); // print out exception to stderr
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
MaybeUnlockForReuse(mtx_shortest, reuse_jbuf_start || reuse_jbuf_limit);
releaseJniEnv(attached_thread);
return;
}
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
if (!reuse_jbuf_start) {
DeleteBuffer(env, j_start_buf);
}
if (!reuse_jbuf_limit) {
DeleteBuffer(env, j_limit_buf);
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
MaybeUnlockForReuse(mtx_shortest, reuse_jbuf_start || reuse_jbuf_limit);
releaseJniEnv(attached_thread);
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
void ComparatorJniCallback::FindShortSuccessor(
std::string* key) const {
if (key == nullptr) {
return;
}
jboolean attached_thread = JNI_FALSE;
JNIEnv* env = getJniEnv(&attached_thread);
assert(env != nullptr);
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
const bool reuse_jbuf_key =
static_cast<int64_t>(key->length()) <= m_options->max_reused_buffer_size;
MaybeLockForReuse(mtx_short, reuse_jbuf_key);
Slice skey(key->data(), key->length());
jobject j_key_buf = GetBuffer(env, skey, reuse_jbuf_key, m_tl_buf_a, m_jshort_buf_key);
if (j_key_buf == nullptr) {
// exception occurred
MaybeUnlockForReuse(mtx_short, reuse_jbuf_key);
env->ExceptionDescribe(); // print out exception to stderr
releaseJniEnv(attached_thread);
return;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
jint jkey_len = env->CallStaticIntMethod(
m_abstract_comparator_jni_bridge_clazz, m_jshort_mid,
m_jcallback_obj,
j_key_buf, reuse_jbuf_key ? key->length() : -1);
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
if (env->ExceptionCheck()) {
// exception thrown from CallObjectMethod
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
if (!reuse_jbuf_key) {
DeleteBuffer(env, j_key_buf);
}
MaybeUnlockForReuse(mtx_short, reuse_jbuf_key);
env->ExceptionDescribe(); // print out exception to stderr
releaseJniEnv(attached_thread);
return;
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
if (static_cast<size_t>(jkey_len) != key->length()) {
// key buffer has changed in Java, so update `key` with the result
bool copy_from_non_direct = false;
if (reuse_jbuf_key) {
// reused a buffer
if (m_options->direct_buffer) {
// reused direct buffer
void* key_buf = env->GetDirectBufferAddress(j_key_buf);
if (key_buf == nullptr) {
rocksdb::RocksDBExceptionJni::ThrowNew(env, "Unable to get Direct Buffer Address");
if (!reuse_jbuf_key) {
DeleteBuffer(env, j_key_buf);
}
MaybeUnlockForReuse(mtx_short, reuse_jbuf_key);
env->ExceptionDescribe(); // print out exception to stderr
releaseJniEnv(attached_thread);
return;
}
key->assign(static_cast<const char*>(key_buf), jkey_len);
} else {
// reused non-direct buffer
copy_from_non_direct = true;
}
} else {
// there was a new buffer
if (m_options->direct_buffer) {
// it was direct... don't forget to potentially truncate the `key` string
key->resize(jkey_len);
} else {
// it was non-direct
copy_from_non_direct = true;
}
}
if (copy_from_non_direct) {
jbyteArray jarray = ByteBufferJni::array(env, j_key_buf,
m_jbytebuffer_clazz);
if (jarray == nullptr) {
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
if (!reuse_jbuf_key) {
DeleteBuffer(env, j_key_buf);
}
MaybeUnlockForReuse(mtx_short, reuse_jbuf_key);
env->ExceptionDescribe(); // print out exception to stderr
releaseJniEnv(attached_thread);
return;
}
jboolean has_exception = JNI_FALSE;
JniUtil::byteString<std::string>(env, jarray, [key, jkey_len](const char* data, const size_t) {
return key->assign(data, static_cast<size_t>(jkey_len));
}, &has_exception);
env->DeleteLocalRef(jarray);
if (has_exception == JNI_TRUE) {
if (!reuse_jbuf_key) {
DeleteBuffer(env, j_key_buf);
}
MaybeUnlockForReuse(mtx_short, reuse_jbuf_key);
env->ExceptionDescribe(); // print out exception to stderr
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
releaseJniEnv(attached_thread);
return;
}
}
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
if (!reuse_jbuf_key) {
DeleteBuffer(env, j_key_buf);
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
MaybeUnlockForReuse(mtx_short, reuse_jbuf_key);
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
releaseJniEnv(attached_thread);
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
inline void ComparatorJniCallback::MaybeLockForReuse(
const std::unique_ptr<port::Mutex>& mutex, const bool cond) const {
// no need to lock if using thread_local
if (m_options->reused_synchronisation_type != ReusedSynchronisationType::THREAD_LOCAL
&& cond) {
mutex.get()->Lock();
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
inline void ComparatorJniCallback::MaybeUnlockForReuse(
const std::unique_ptr<port::Mutex>& mutex, const bool cond) const {
// no need to unlock if using thread_local
if (m_options->reused_synchronisation_type != ReusedSynchronisationType::THREAD_LOCAL
&& cond) {
mutex.get()->Unlock();
}
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
jobject ComparatorJniCallback::GetBuffer(JNIEnv* env, const Slice& src,
bool reuse_buffer, ThreadLocalPtr* tl_buf, jobject jreuse_buffer) const {
if (reuse_buffer) {
if (m_options->reused_synchronisation_type
== ReusedSynchronisationType::THREAD_LOCAL) {
// reuse thread-local bufffer
ThreadLocalBuf* tlb = reinterpret_cast<ThreadLocalBuf*>(tl_buf->Get());
if (tlb == nullptr) {
// thread-local buffer has not yet been created, so create it
jobject jtl_buf = env->NewGlobalRef(ByteBufferJni::construct(
env, m_options->direct_buffer, m_options->max_reused_buffer_size,
m_jbytebuffer_clazz));
if (jtl_buf == nullptr) {
// exception thrown: OutOfMemoryError
return nullptr;
}
tlb = new ThreadLocalBuf(m_jvm, m_options->direct_buffer, jtl_buf);
tl_buf->Reset(tlb);
}
return ReuseBuffer(env, src, tlb->jbuf);
} else {
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
// reuse class member buffer
return ReuseBuffer(env, src, jreuse_buffer);
}
} else {
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
// new buffer
return NewBuffer(env, src);
}
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
jobject ComparatorJniCallback::ReuseBuffer(
JNIEnv* env, const Slice& src, jobject jreuse_buffer) const {
// we can reuse the buffer
if (m_options->direct_buffer) {
// copy into direct buffer
void* buf = env->GetDirectBufferAddress(jreuse_buffer);
if (buf == nullptr) {
// either memory region is undefined, given object is not a direct java.nio.Buffer, or JNI access to direct buffers is not supported by this virtual machine.
rocksdb::RocksDBExceptionJni::ThrowNew(env, "Unable to get Direct Buffer Address");
return nullptr;
}
memcpy(buf, src.data(), src.size());
} else {
// copy into non-direct buffer
const jbyteArray jarray = ByteBufferJni::array(env, jreuse_buffer,
m_jbytebuffer_clazz);
if (jarray == nullptr) {
// exception occurred
return nullptr;
}
env->SetByteArrayRegion(jarray, 0, static_cast<jsize>(src.size()),
const_cast<jbyte*>(reinterpret_cast<const jbyte*>(src.data())));
if (env->ExceptionCheck()) {
// exception occurred
env->DeleteLocalRef(jarray);
return nullptr;
}
env->DeleteLocalRef(jarray);
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
return jreuse_buffer;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
jobject ComparatorJniCallback::NewBuffer(JNIEnv* env, const Slice& src) const {
// we need a new buffer
jobject jbuf = ByteBufferJni::constructWith(env, m_options->direct_buffer,
src.data(), src.size(), m_jbytebuffer_clazz);
if (jbuf == nullptr) {
// exception occurred
return nullptr;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
return jbuf;
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
void ComparatorJniCallback::DeleteBuffer(JNIEnv* env, jobject jbuffer) const {
env->DeleteLocalRef(jbuffer);
}
Improve RocksJava Comparator (#6252) Summary: This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy. **NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into. Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`). In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`. In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`. --- [JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java. With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425). These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical. --- These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x. ``` ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252 Differential Revision: D19331064 Pulled By: pdillinger fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
} // namespace rocksdb