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/cache/cache_test.cc

960 lines
29 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).
//
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.
#include "rocksdb/cache.h"
#include <forward_list>
#include <functional>
#include <iostream>
#include <string>
#include <vector>
#include "cache/clock_cache.h"
#include "cache/fast_lru_cache.h"
#include "cache/lru_cache.h"
#include "port/stack_trace.h"
#include "test_util/testharness.h"
#include "util/coding.h"
#include "util/string_util.h"
namespace ROCKSDB_NAMESPACE {
namespace {
// Conversions between numeric keys/values and the types expected by Cache.
std::string EncodeKey16Bytes(int k) {
std::string result;
PutFixed32(&result, k);
result.append(std::string(12, 'a')); // Because we need a 16B output, we
// add a 12-byte padding.
return result;
}
int DecodeKey16Bytes(const Slice& k) {
assert(k.size() == 16);
return DecodeFixed32(k.data()); // Decodes only the first 4 bytes of k.
}
std::string EncodeKey32Bits(int k) {
std::string result;
PutFixed32(&result, k);
return result;
}
int DecodeKey32Bits(const Slice& k) {
assert(k.size() == 4);
return DecodeFixed32(k.data());
}
void* EncodeValue(uintptr_t v) { return reinterpret_cast<void*>(v); }
int DecodeValue(void* v) {
return static_cast<int>(reinterpret_cast<uintptr_t>(v));
}
const std::string kLRU = "lru";
const std::string kClock = "clock";
const std::string kFast = "fast";
void dumbDeleter(const Slice& /*key*/, void* /*value*/) {}
void eraseDeleter(const Slice& /*key*/, void* value) {
Cache* cache = reinterpret_cast<Cache*>(value);
cache->Erase("foo");
}
} // anonymous namespace
class CacheTest : public testing::TestWithParam<std::string> {
public:
static CacheTest* current_;
static std::string type_;
static void Deleter(const Slice& key, void* v) {
if (type_ == kFast) {
current_->deleted_keys_.push_back(DecodeKey16Bytes(key));
} else {
current_->deleted_keys_.push_back(DecodeKey32Bits(key));
}
current_->deleted_values_.push_back(DecodeValue(v));
}
static const int kCacheSize = 1000;
static const int kNumShardBits = 4;
static const int kCacheSize2 = 100;
static const int kNumShardBits2 = 2;
std::vector<int> deleted_keys_;
std::vector<int> deleted_values_;
std::shared_ptr<Cache> cache_;
std::shared_ptr<Cache> cache2_;
CacheTest()
: cache_(NewCache(kCacheSize, kNumShardBits, false)),
cache2_(NewCache(kCacheSize2, kNumShardBits2, false)) {
current_ = this;
type_ = GetParam();
}
~CacheTest() override {}
std::shared_ptr<Cache> NewCache(size_t capacity) {
auto type = GetParam();
if (type == kLRU) {
return NewLRUCache(capacity);
}
if (type == kClock) {
return NewClockCache(capacity);
}
if (type == kFast) {
return NewFastLRUCache(
capacity, 1 /*estimated_value_size*/, -1 /*num_shard_bits*/,
false /*strict_capacity_limit*/, kDefaultCacheMetadataChargePolicy);
}
return nullptr;
}
std::shared_ptr<Cache> NewCache(
size_t capacity, int num_shard_bits, bool strict_capacity_limit,
CacheMetadataChargePolicy charge_policy = kDontChargeCacheMetadata) {
auto type = GetParam();
if (type == kLRU) {
LRUCacheOptions co;
co.capacity = capacity;
co.num_shard_bits = num_shard_bits;
co.strict_capacity_limit = strict_capacity_limit;
co.high_pri_pool_ratio = 0;
co.metadata_charge_policy = charge_policy;
return NewLRUCache(co);
}
if (type == kClock) {
return NewClockCache(capacity, num_shard_bits, strict_capacity_limit,
charge_policy);
}
if (type == kFast) {
return NewFastLRUCache(capacity, 1 /*estimated_value_size*/,
num_shard_bits, strict_capacity_limit,
charge_policy);
}
return nullptr;
}
// These functions encode/decode keys in tests cases that use
// int keys.
// Currently, FastLRUCache requires keys to be 16B long, whereas
// LRUCache and ClockCache don't, so the encoding depends on
// the cache type.
std::string EncodeKey(int k) {
if (GetParam() == kFast) {
return EncodeKey16Bytes(k);
} else {
return EncodeKey32Bits(k);
}
}
int DecodeKey(const Slice& k) {
if (GetParam() == kFast) {
return DecodeKey16Bytes(k);
} else {
return DecodeKey32Bits(k);
}
}
int Lookup(std::shared_ptr<Cache> cache, int key) {
Cache::Handle* handle = cache->Lookup(EncodeKey(key));
const int r = (handle == nullptr) ? -1 : DecodeValue(cache->Value(handle));
if (handle != nullptr) {
cache->Release(handle);
}
return r;
}
void Insert(std::shared_ptr<Cache> cache, int key, int value,
int charge = 1) {
EXPECT_OK(cache->Insert(EncodeKey(key), EncodeValue(value), charge,
&CacheTest::Deleter));
}
void Erase(std::shared_ptr<Cache> cache, int key) {
cache->Erase(EncodeKey(key));
}
int Lookup(int key) {
return Lookup(cache_, key);
}
void Insert(int key, int value, int charge = 1) {
Insert(cache_, key, value, charge);
}
void Erase(int key) {
Erase(cache_, key);
}
int Lookup2(int key) {
return Lookup(cache2_, key);
}
void Insert2(int key, int value, int charge = 1) {
Insert(cache2_, key, value, charge);
}
void Erase2(int key) {
Erase(cache2_, key);
}
};
CacheTest* CacheTest::current_;
std::string CacheTest::type_;
class LRUCacheTest : public CacheTest {};
TEST_P(CacheTest, UsageTest) {
if (GetParam() == kFast) {
ROCKSDB_GTEST_BYPASS("FastLRUCache requires 16 byte keys.");
return;
}
// cache is std::shared_ptr and will be automatically cleaned up.
const uint64_t kCapacity = 100000;
auto cache = NewCache(kCapacity, 8, false, kDontChargeCacheMetadata);
auto precise_cache = NewCache(kCapacity, 0, false, kFullChargeCacheMetadata);
ASSERT_EQ(0, cache->GetUsage());
ASSERT_EQ(0, precise_cache->GetUsage());
size_t usage = 0;
char value[10] = "abcdef";
// make sure everything will be cached
for (int i = 1; i < 100; ++i) {
std::string key(i, 'a');
auto kv_size = key.size() + 5;
ASSERT_OK(cache->Insert(key, reinterpret_cast<void*>(value), kv_size,
dumbDeleter));
ASSERT_OK(precise_cache->Insert(key, reinterpret_cast<void*>(value),
kv_size, dumbDeleter));
usage += kv_size;
ASSERT_EQ(usage, cache->GetUsage());
ASSERT_LT(usage, precise_cache->GetUsage());
}
cache->EraseUnRefEntries();
precise_cache->EraseUnRefEntries();
ASSERT_EQ(0, cache->GetUsage());
ASSERT_EQ(0, precise_cache->GetUsage());
// make sure the cache will be overloaded
for (uint64_t i = 1; i < kCapacity; ++i) {
auto key = std::to_string(i);
ASSERT_OK(cache->Insert(key, reinterpret_cast<void*>(value), key.size() + 5,
dumbDeleter));
ASSERT_OK(precise_cache->Insert(key, reinterpret_cast<void*>(value),
key.size() + 5, dumbDeleter));
}
// the usage should be close to the capacity
ASSERT_GT(kCapacity, cache->GetUsage());
ASSERT_GT(kCapacity, precise_cache->GetUsage());
ASSERT_LT(kCapacity * 0.95, cache->GetUsage());
ASSERT_LT(kCapacity * 0.95, precise_cache->GetUsage());
}
TEST_P(CacheTest, PinnedUsageTest) {
if (GetParam() == kFast) {
ROCKSDB_GTEST_BYPASS("FastLRUCache requires 16 byte keys.");
return;
}
// cache is std::shared_ptr and will be automatically cleaned up.
const uint64_t kCapacity = 200000;
auto cache = NewCache(kCapacity, 8, false, kDontChargeCacheMetadata);
auto precise_cache = NewCache(kCapacity, 8, false, kFullChargeCacheMetadata);
size_t pinned_usage = 0;
char value[10] = "abcdef";
std::forward_list<Cache::Handle*> unreleased_handles;
std::forward_list<Cache::Handle*> unreleased_handles_in_precise_cache;
// Add entries. Unpin some of them after insertion. Then, pin some of them
// again. Check GetPinnedUsage().
for (int i = 1; i < 100; ++i) {
std::string key(i, 'a');
auto kv_size = key.size() + 5;
Cache::Handle* handle;
Cache::Handle* handle_in_precise_cache;
ASSERT_OK(cache->Insert(key, reinterpret_cast<void*>(value), kv_size,
dumbDeleter, &handle));
assert(handle);
ASSERT_OK(precise_cache->Insert(key, reinterpret_cast<void*>(value),
kv_size, dumbDeleter,
&handle_in_precise_cache));
assert(handle_in_precise_cache);
pinned_usage += kv_size;
ASSERT_EQ(pinned_usage, cache->GetPinnedUsage());
ASSERT_LT(pinned_usage, precise_cache->GetPinnedUsage());
if (i % 2 == 0) {
cache->Release(handle);
precise_cache->Release(handle_in_precise_cache);
pinned_usage -= kv_size;
ASSERT_EQ(pinned_usage, cache->GetPinnedUsage());
ASSERT_LT(pinned_usage, precise_cache->GetPinnedUsage());
} else {
unreleased_handles.push_front(handle);
unreleased_handles_in_precise_cache.push_front(handle_in_precise_cache);
}
if (i % 3 == 0) {
unreleased_handles.push_front(cache->Lookup(key));
auto x = precise_cache->Lookup(key);
assert(x);
unreleased_handles_in_precise_cache.push_front(x);
// If i % 2 == 0, then the entry was unpinned before Lookup, so pinned
// usage increased
if (i % 2 == 0) {
pinned_usage += kv_size;
}
ASSERT_EQ(pinned_usage, cache->GetPinnedUsage());
ASSERT_LT(pinned_usage, precise_cache->GetPinnedUsage());
}
}
auto precise_cache_pinned_usage = precise_cache->GetPinnedUsage();
ASSERT_LT(pinned_usage, precise_cache_pinned_usage);
// check that overloading the cache does not change the pinned usage
for (uint64_t i = 1; i < 2 * kCapacity; ++i) {
auto key = std::to_string(i);
ASSERT_OK(cache->Insert(key, reinterpret_cast<void*>(value), key.size() + 5,
dumbDeleter));
ASSERT_OK(precise_cache->Insert(key, reinterpret_cast<void*>(value),
key.size() + 5, dumbDeleter));
}
ASSERT_EQ(pinned_usage, cache->GetPinnedUsage());
ASSERT_EQ(precise_cache_pinned_usage, precise_cache->GetPinnedUsage());
cache->EraseUnRefEntries();
precise_cache->EraseUnRefEntries();
ASSERT_EQ(pinned_usage, cache->GetPinnedUsage());
ASSERT_EQ(precise_cache_pinned_usage, precise_cache->GetPinnedUsage());
// release handles for pinned entries to prevent memory leaks
for (auto handle : unreleased_handles) {
cache->Release(handle);
}
for (auto handle : unreleased_handles_in_precise_cache) {
precise_cache->Release(handle);
}
ASSERT_EQ(0, cache->GetPinnedUsage());
ASSERT_EQ(0, precise_cache->GetPinnedUsage());
cache->EraseUnRefEntries();
precise_cache->EraseUnRefEntries();
ASSERT_EQ(0, cache->GetUsage());
ASSERT_EQ(0, precise_cache->GetUsage());
}
TEST_P(CacheTest, HitAndMiss) {
ASSERT_EQ(-1, Lookup(100));
Insert(100, 101);
ASSERT_EQ(101, Lookup(100));
ASSERT_EQ(-1, Lookup(200));
ASSERT_EQ(-1, Lookup(300));
Insert(200, 201);
ASSERT_EQ(101, Lookup(100));
ASSERT_EQ(201, Lookup(200));
ASSERT_EQ(-1, Lookup(300));
Insert(100, 102);
ASSERT_EQ(102, Lookup(100));
ASSERT_EQ(201, Lookup(200));
ASSERT_EQ(-1, Lookup(300));
ASSERT_EQ(1U, deleted_keys_.size());
ASSERT_EQ(100, deleted_keys_[0]);
ASSERT_EQ(101, deleted_values_[0]);
}
TEST_P(CacheTest, InsertSameKey) {
Insert(1, 1);
Insert(1, 2);
ASSERT_EQ(2, Lookup(1));
}
TEST_P(CacheTest, Erase) {
Erase(200);
ASSERT_EQ(0U, deleted_keys_.size());
Insert(100, 101);
Insert(200, 201);
Erase(100);
ASSERT_EQ(-1, Lookup(100));
ASSERT_EQ(201, Lookup(200));
ASSERT_EQ(1U, deleted_keys_.size());
ASSERT_EQ(100, deleted_keys_[0]);
ASSERT_EQ(101, deleted_values_[0]);
Erase(100);
ASSERT_EQ(-1, Lookup(100));
ASSERT_EQ(201, Lookup(200));
ASSERT_EQ(1U, deleted_keys_.size());
}
TEST_P(CacheTest, EntriesArePinned) {
Insert(100, 101);
Cache::Handle* h1 = cache_->Lookup(EncodeKey(100));
ASSERT_EQ(101, DecodeValue(cache_->Value(h1)));
ASSERT_EQ(1U, cache_->GetUsage());
Insert(100, 102);
Cache::Handle* h2 = cache_->Lookup(EncodeKey(100));
ASSERT_EQ(102, DecodeValue(cache_->Value(h2)));
ASSERT_EQ(0U, deleted_keys_.size());
ASSERT_EQ(2U, cache_->GetUsage());
cache_->Release(h1);
ASSERT_EQ(1U, deleted_keys_.size());
ASSERT_EQ(100, deleted_keys_[0]);
ASSERT_EQ(101, deleted_values_[0]);
ASSERT_EQ(1U, cache_->GetUsage());
Erase(100);
ASSERT_EQ(-1, Lookup(100));
ASSERT_EQ(1U, deleted_keys_.size());
ASSERT_EQ(1U, cache_->GetUsage());
cache_->Release(h2);
ASSERT_EQ(2U, deleted_keys_.size());
ASSERT_EQ(100, deleted_keys_[1]);
ASSERT_EQ(102, deleted_values_[1]);
ASSERT_EQ(0U, cache_->GetUsage());
}
TEST_P(CacheTest, EvictionPolicy) {
Insert(100, 101);
Insert(200, 201);
// Frequently used entry must be kept around
for (int i = 0; i < kCacheSize * 2; i++) {
Insert(1000+i, 2000+i);
ASSERT_EQ(101, Lookup(100));
}
ASSERT_EQ(101, Lookup(100));
ASSERT_EQ(-1, Lookup(200));
}
TEST_P(CacheTest, ExternalRefPinsEntries) {
Insert(100, 101);
Cache::Handle* h = cache_->Lookup(EncodeKey(100));
ASSERT_TRUE(cache_->Ref(h));
ASSERT_EQ(101, DecodeValue(cache_->Value(h)));
ASSERT_EQ(1U, cache_->GetUsage());
for (int i = 0; i < 3; ++i) {
if (i > 0) {
// First release (i == 1) corresponds to Ref(), second release (i == 2)
// corresponds to Lookup(). Then, since all external refs are released,
// the below insertions should push out the cache entry.
cache_->Release(h);
}
// double cache size because the usage bit in block cache prevents 100 from
// being evicted in the first kCacheSize iterations
for (int j = 0; j < 2 * kCacheSize + 100; j++) {
Insert(1000 + j, 2000 + j);
}
if (i < 2) {
ASSERT_EQ(101, Lookup(100));
}
}
ASSERT_EQ(-1, Lookup(100));
}
TEST_P(CacheTest, EvictionPolicyRef) {
Insert(100, 101);
Insert(101, 102);
Insert(102, 103);
Insert(103, 104);
Insert(200, 101);
Insert(201, 102);
Insert(202, 103);
Insert(203, 104);
Cache::Handle* h201 = cache_->Lookup(EncodeKey(200));
Cache::Handle* h202 = cache_->Lookup(EncodeKey(201));
Cache::Handle* h203 = cache_->Lookup(EncodeKey(202));
Cache::Handle* h204 = cache_->Lookup(EncodeKey(203));
Insert(300, 101);
Insert(301, 102);
Insert(302, 103);
Insert(303, 104);
// Insert entries much more than Cache capacity
for (int i = 0; i < kCacheSize * 2; i++) {
Insert(1000 + i, 2000 + i);
}
// Check whether the entries inserted in the beginning
// are evicted. Ones without extra ref are evicted and
// those with are not.
ASSERT_EQ(-1, Lookup(100));
ASSERT_EQ(-1, Lookup(101));
ASSERT_EQ(-1, Lookup(102));
ASSERT_EQ(-1, Lookup(103));
ASSERT_EQ(-1, Lookup(300));
ASSERT_EQ(-1, Lookup(301));
ASSERT_EQ(-1, Lookup(302));
ASSERT_EQ(-1, Lookup(303));
ASSERT_EQ(101, Lookup(200));
ASSERT_EQ(102, Lookup(201));
ASSERT_EQ(103, Lookup(202));
ASSERT_EQ(104, Lookup(203));
// Cleaning up all the handles
cache_->Release(h201);
cache_->Release(h202);
cache_->Release(h203);
cache_->Release(h204);
}
TEST_P(CacheTest, EvictEmptyCache) {
if (GetParam() == kFast) {
ROCKSDB_GTEST_BYPASS("FastLRUCache requires 16 byte keys.");
return;
}
// Insert item large than capacity to trigger eviction on empty cache.
auto cache = NewCache(1, 0, false);
ASSERT_OK(cache->Insert("foo", nullptr, 10, dumbDeleter));
}
TEST_P(CacheTest, EraseFromDeleter) {
if (GetParam() == kFast) {
ROCKSDB_GTEST_BYPASS("FastLRUCache requires 16 byte keys.");
return;
}
// Have deleter which will erase item from cache, which will re-enter
// the cache at that point.
std::shared_ptr<Cache> cache = NewCache(10, 0, false);
ASSERT_OK(cache->Insert("foo", nullptr, 1, dumbDeleter));
ASSERT_OK(cache->Insert("bar", cache.get(), 1, eraseDeleter));
cache->Erase("bar");
ASSERT_EQ(nullptr, cache->Lookup("foo"));
ASSERT_EQ(nullptr, cache->Lookup("bar"));
}
TEST_P(CacheTest, ErasedHandleState) {
Modifed the LRU cache eviction code so that it doesn't evict blocks which have exteranl references Summary: Currently, blocks which have more than one reference (ie referenced by something other than cache itself) are evicted from cache. This doesn't make much sense: - blocks are still in RAM, so the RAM usage reported by the cache is incorrect - if the same block is needed by another iterator, it will be loaded and decompressed again This diff changes the reference counting scheme a bit. Previously, if the cache contained the block, this was accounted for in its refcount. After this change, the refcount is only used to track external references. There is a boolean flag which indicates whether or not the block is contained in the cache. This diff also changes how LRU list is used. Previously, both hashtable and the LRU list contained all blocks. After this change, the LRU list contains blocks with the refcount==0, ie those which can be evicted from the cache. Note that this change still allows for cache to grow beyond its capacity. This happens when all blocks are pinned (ie refcount>0). This is consistent with the current behavior. The cache's insert function never fails. I spent lots of time trying to make table_reader and other places work with the insert which might failed. It turned out to be pretty hard. It might really destabilize some customers, so finally, I decided against doing this. table_cache_remove_scan_count_limit option will be unneeded after this change, but I will remove it in the following diff, if this one gets approved Test Plan: Ran tests, made sure they pass Reviewers: sdong, ljin Differential Revision: https://reviews.facebook.net/D25503
10 years ago
// insert a key and get two handles
Insert(100, 1000);
Cache::Handle* h1 = cache_->Lookup(EncodeKey(100));
Cache::Handle* h2 = cache_->Lookup(EncodeKey(100));
ASSERT_EQ(h1, h2);
ASSERT_EQ(DecodeValue(cache_->Value(h1)), 1000);
ASSERT_EQ(DecodeValue(cache_->Value(h2)), 1000);
Modifed the LRU cache eviction code so that it doesn't evict blocks which have exteranl references Summary: Currently, blocks which have more than one reference (ie referenced by something other than cache itself) are evicted from cache. This doesn't make much sense: - blocks are still in RAM, so the RAM usage reported by the cache is incorrect - if the same block is needed by another iterator, it will be loaded and decompressed again This diff changes the reference counting scheme a bit. Previously, if the cache contained the block, this was accounted for in its refcount. After this change, the refcount is only used to track external references. There is a boolean flag which indicates whether or not the block is contained in the cache. This diff also changes how LRU list is used. Previously, both hashtable and the LRU list contained all blocks. After this change, the LRU list contains blocks with the refcount==0, ie those which can be evicted from the cache. Note that this change still allows for cache to grow beyond its capacity. This happens when all blocks are pinned (ie refcount>0). This is consistent with the current behavior. The cache's insert function never fails. I spent lots of time trying to make table_reader and other places work with the insert which might failed. It turned out to be pretty hard. It might really destabilize some customers, so finally, I decided against doing this. table_cache_remove_scan_count_limit option will be unneeded after this change, but I will remove it in the following diff, if this one gets approved Test Plan: Ran tests, made sure they pass Reviewers: sdong, ljin Differential Revision: https://reviews.facebook.net/D25503
10 years ago
// delete the key from the cache
Erase(100);
// can no longer find in the cache
ASSERT_EQ(-1, Lookup(100));
Modifed the LRU cache eviction code so that it doesn't evict blocks which have exteranl references Summary: Currently, blocks which have more than one reference (ie referenced by something other than cache itself) are evicted from cache. This doesn't make much sense: - blocks are still in RAM, so the RAM usage reported by the cache is incorrect - if the same block is needed by another iterator, it will be loaded and decompressed again This diff changes the reference counting scheme a bit. Previously, if the cache contained the block, this was accounted for in its refcount. After this change, the refcount is only used to track external references. There is a boolean flag which indicates whether or not the block is contained in the cache. This diff also changes how LRU list is used. Previously, both hashtable and the LRU list contained all blocks. After this change, the LRU list contains blocks with the refcount==0, ie those which can be evicted from the cache. Note that this change still allows for cache to grow beyond its capacity. This happens when all blocks are pinned (ie refcount>0). This is consistent with the current behavior. The cache's insert function never fails. I spent lots of time trying to make table_reader and other places work with the insert which might failed. It turned out to be pretty hard. It might really destabilize some customers, so finally, I decided against doing this. table_cache_remove_scan_count_limit option will be unneeded after this change, but I will remove it in the following diff, if this one gets approved Test Plan: Ran tests, made sure they pass Reviewers: sdong, ljin Differential Revision: https://reviews.facebook.net/D25503
10 years ago
// release one handle
cache_->Release(h1);
// still can't find in cache
ASSERT_EQ(-1, Lookup(100));
Modifed the LRU cache eviction code so that it doesn't evict blocks which have exteranl references Summary: Currently, blocks which have more than one reference (ie referenced by something other than cache itself) are evicted from cache. This doesn't make much sense: - blocks are still in RAM, so the RAM usage reported by the cache is incorrect - if the same block is needed by another iterator, it will be loaded and decompressed again This diff changes the reference counting scheme a bit. Previously, if the cache contained the block, this was accounted for in its refcount. After this change, the refcount is only used to track external references. There is a boolean flag which indicates whether or not the block is contained in the cache. This diff also changes how LRU list is used. Previously, both hashtable and the LRU list contained all blocks. After this change, the LRU list contains blocks with the refcount==0, ie those which can be evicted from the cache. Note that this change still allows for cache to grow beyond its capacity. This happens when all blocks are pinned (ie refcount>0). This is consistent with the current behavior. The cache's insert function never fails. I spent lots of time trying to make table_reader and other places work with the insert which might failed. It turned out to be pretty hard. It might really destabilize some customers, so finally, I decided against doing this. table_cache_remove_scan_count_limit option will be unneeded after this change, but I will remove it in the following diff, if this one gets approved Test Plan: Ran tests, made sure they pass Reviewers: sdong, ljin Differential Revision: https://reviews.facebook.net/D25503
10 years ago
cache_->Release(h2);
}
TEST_P(CacheTest, HeavyEntries) {
// Add a bunch of light and heavy entries and then count the combined
// size of items still in the cache, which must be approximately the
// same as the total capacity.
const int kLight = 1;
const int kHeavy = 10;
int added = 0;
int index = 0;
while (added < 2*kCacheSize) {
const int weight = (index & 1) ? kLight : kHeavy;
Insert(index, 1000+index, weight);
added += weight;
index++;
}
int cached_weight = 0;
for (int i = 0; i < index; i++) {
const int weight = (i & 1 ? kLight : kHeavy);
int r = Lookup(i);
if (r >= 0) {
cached_weight += weight;
ASSERT_EQ(1000+i, r);
}
}
ASSERT_LE(cached_weight, kCacheSize + kCacheSize/10);
}
TEST_P(CacheTest, NewId) {
uint64_t a = cache_->NewId();
uint64_t b = cache_->NewId();
ASSERT_NE(a, b);
}
class Value {
public:
explicit Value(int v) : v_(v) {}
int v_;
};
namespace {
void deleter(const Slice& /*key*/, void* value) {
delete static_cast<Value *>(value);
}
} // namespace
TEST_P(CacheTest, ReleaseAndErase) {
std::shared_ptr<Cache> cache = NewCache(5, 0, false);
Cache::Handle* handle;
Status s = cache->Insert(EncodeKey(100), EncodeValue(100), 1,
&CacheTest::Deleter, &handle);
ASSERT_TRUE(s.ok());
ASSERT_EQ(5U, cache->GetCapacity());
ASSERT_EQ(1U, cache->GetUsage());
ASSERT_EQ(0U, deleted_keys_.size());
auto erased = cache->Release(handle, true);
ASSERT_TRUE(erased);
// This tests that deleter has been called
ASSERT_EQ(1U, deleted_keys_.size());
}
TEST_P(CacheTest, ReleaseWithoutErase) {
std::shared_ptr<Cache> cache = NewCache(5, 0, false);
Cache::Handle* handle;
Status s = cache->Insert(EncodeKey(100), EncodeValue(100), 1,
&CacheTest::Deleter, &handle);
ASSERT_TRUE(s.ok());
ASSERT_EQ(5U, cache->GetCapacity());
ASSERT_EQ(1U, cache->GetUsage());
ASSERT_EQ(0U, deleted_keys_.size());
auto erased = cache->Release(handle);
ASSERT_FALSE(erased);
// This tests that deleter is not called. When cache has free capacity it is
// not expected to immediately erase the released items.
ASSERT_EQ(0U, deleted_keys_.size());
}
TEST_P(CacheTest, SetCapacity) {
if (GetParam() == kFast) {
ROCKSDB_GTEST_BYPASS("FastLRUCache doesn't support capacity adjustments.");
return;
}
// test1: increase capacity
// lets create a cache with capacity 5,
// then, insert 5 elements, then increase capacity
// to 10, returned capacity should be 10, usage=5
std::shared_ptr<Cache> cache = NewCache(5, 0, false);
std::vector<Cache::Handle*> handles(10);
// Insert 5 entries, but not releasing.
for (int i = 0; i < 5; i++) {
std::string key = EncodeKey(i + 1);
Status s = cache->Insert(key, new Value(i + 1), 1, &deleter, &handles[i]);
ASSERT_TRUE(s.ok());
}
ASSERT_EQ(5U, cache->GetCapacity());
ASSERT_EQ(5U, cache->GetUsage());
cache->SetCapacity(10);
ASSERT_EQ(10U, cache->GetCapacity());
ASSERT_EQ(5U, cache->GetUsage());
// test2: decrease capacity
// insert 5 more elements to cache, then release 5,
// then decrease capacity to 7, final capacity should be 7
// and usage should be 7
for (int i = 5; i < 10; i++) {
std::string key = EncodeKey(i + 1);
Status s = cache->Insert(key, new Value(i + 1), 1, &deleter, &handles[i]);
ASSERT_TRUE(s.ok());
}
ASSERT_EQ(10U, cache->GetCapacity());
ASSERT_EQ(10U, cache->GetUsage());
for (int i = 0; i < 5; i++) {
cache->Release(handles[i]);
}
ASSERT_EQ(10U, cache->GetCapacity());
ASSERT_EQ(10U, cache->GetUsage());
cache->SetCapacity(7);
ASSERT_EQ(7, cache->GetCapacity());
ASSERT_EQ(7, cache->GetUsage());
// release remaining 5 to keep valgrind happy
for (int i = 5; i < 10; i++) {
cache->Release(handles[i]);
}
// Make sure this doesn't crash or upset ASAN/valgrind
cache->DisownData();
}
TEST_P(LRUCacheTest, SetStrictCapacityLimit) {
if (GetParam() == kFast) {
ROCKSDB_GTEST_BYPASS(
"FastLRUCache doesn't support an unbounded number of inserts beyond "
"capacity.");
return;
}
// test1: set the flag to false. Insert more keys than capacity. See if they
// all go through.
std::shared_ptr<Cache> cache = NewCache(5, 0, false);
std::vector<Cache::Handle*> handles(10);
Status s;
for (int i = 0; i < 10; i++) {
std::string key = EncodeKey(i + 1);
s = cache->Insert(key, new Value(i + 1), 1, &deleter, &handles[i]);
ASSERT_OK(s);
ASSERT_NE(nullptr, handles[i]);
}
ASSERT_EQ(10, cache->GetUsage());
// test2: set the flag to true. Insert and check if it fails.
std::string extra_key = EncodeKey(100);
Value* extra_value = new Value(0);
cache->SetStrictCapacityLimit(true);
Cache::Handle* handle;
s = cache->Insert(extra_key, extra_value, 1, &deleter, &handle);
ASSERT_TRUE(s.IsIncomplete());
ASSERT_EQ(nullptr, handle);
ASSERT_EQ(10, cache->GetUsage());
for (int i = 0; i < 10; i++) {
cache->Release(handles[i]);
}
// test3: init with flag being true.
std::shared_ptr<Cache> cache2 = NewCache(5, 0, true);
for (int i = 0; i < 5; i++) {
std::string key = EncodeKey(i + 1);
s = cache2->Insert(key, new Value(i + 1), 1, &deleter, &handles[i]);
ASSERT_OK(s);
ASSERT_NE(nullptr, handles[i]);
}
s = cache2->Insert(extra_key, extra_value, 1, &deleter, &handle);
ASSERT_TRUE(s.IsIncomplete());
ASSERT_EQ(nullptr, handle);
// test insert without handle
s = cache2->Insert(extra_key, extra_value, 1, &deleter);
// AS if the key have been inserted into cache but get evicted immediately.
ASSERT_OK(s);
ASSERT_EQ(5, cache2->GetUsage());
ASSERT_EQ(nullptr, cache2->Lookup(extra_key));
for (int i = 0; i < 5; i++) {
cache2->Release(handles[i]);
}
}
TEST_P(CacheTest, OverCapacity) {
if (GetParam() == kFast) {
ROCKSDB_GTEST_BYPASS("FastLRUCache doesn't support capacity adjustments.");
return;
}
size_t n = 10;
// a LRUCache with n entries and one shard only
std::shared_ptr<Cache> cache = NewCache(n, 0, false);
std::vector<Cache::Handle*> handles(n+1);
// Insert n+1 entries, but not releasing.
for (int i = 0; i < static_cast<int>(n + 1); i++) {
std::string key = EncodeKey(i + 1);
Status s = cache->Insert(key, new Value(i + 1), 1, &deleter, &handles[i]);
ASSERT_TRUE(s.ok());
}
// Guess what's in the cache now?
for (int i = 0; i < static_cast<int>(n + 1); i++) {
std::string key = EncodeKey(i + 1);
auto h = cache->Lookup(key);
Modifed the LRU cache eviction code so that it doesn't evict blocks which have exteranl references Summary: Currently, blocks which have more than one reference (ie referenced by something other than cache itself) are evicted from cache. This doesn't make much sense: - blocks are still in RAM, so the RAM usage reported by the cache is incorrect - if the same block is needed by another iterator, it will be loaded and decompressed again This diff changes the reference counting scheme a bit. Previously, if the cache contained the block, this was accounted for in its refcount. After this change, the refcount is only used to track external references. There is a boolean flag which indicates whether or not the block is contained in the cache. This diff also changes how LRU list is used. Previously, both hashtable and the LRU list contained all blocks. After this change, the LRU list contains blocks with the refcount==0, ie those which can be evicted from the cache. Note that this change still allows for cache to grow beyond its capacity. This happens when all blocks are pinned (ie refcount>0). This is consistent with the current behavior. The cache's insert function never fails. I spent lots of time trying to make table_reader and other places work with the insert which might failed. It turned out to be pretty hard. It might really destabilize some customers, so finally, I decided against doing this. table_cache_remove_scan_count_limit option will be unneeded after this change, but I will remove it in the following diff, if this one gets approved Test Plan: Ran tests, made sure they pass Reviewers: sdong, ljin Differential Revision: https://reviews.facebook.net/D25503
10 years ago
ASSERT_TRUE(h != nullptr);
if (h) cache->Release(h);
}
Modifed the LRU cache eviction code so that it doesn't evict blocks which have exteranl references Summary: Currently, blocks which have more than one reference (ie referenced by something other than cache itself) are evicted from cache. This doesn't make much sense: - blocks are still in RAM, so the RAM usage reported by the cache is incorrect - if the same block is needed by another iterator, it will be loaded and decompressed again This diff changes the reference counting scheme a bit. Previously, if the cache contained the block, this was accounted for in its refcount. After this change, the refcount is only used to track external references. There is a boolean flag which indicates whether or not the block is contained in the cache. This diff also changes how LRU list is used. Previously, both hashtable and the LRU list contained all blocks. After this change, the LRU list contains blocks with the refcount==0, ie those which can be evicted from the cache. Note that this change still allows for cache to grow beyond its capacity. This happens when all blocks are pinned (ie refcount>0). This is consistent with the current behavior. The cache's insert function never fails. I spent lots of time trying to make table_reader and other places work with the insert which might failed. It turned out to be pretty hard. It might really destabilize some customers, so finally, I decided against doing this. table_cache_remove_scan_count_limit option will be unneeded after this change, but I will remove it in the following diff, if this one gets approved Test Plan: Ran tests, made sure they pass Reviewers: sdong, ljin Differential Revision: https://reviews.facebook.net/D25503
10 years ago
// the cache is over capacity since nothing could be evicted
ASSERT_EQ(n + 1U, cache->GetUsage());
for (int i = 0; i < static_cast<int>(n + 1); i++) {
cache->Release(handles[i]);
}
// Make sure eviction is triggered.
cache->SetCapacity(n);
Modifed the LRU cache eviction code so that it doesn't evict blocks which have exteranl references Summary: Currently, blocks which have more than one reference (ie referenced by something other than cache itself) are evicted from cache. This doesn't make much sense: - blocks are still in RAM, so the RAM usage reported by the cache is incorrect - if the same block is needed by another iterator, it will be loaded and decompressed again This diff changes the reference counting scheme a bit. Previously, if the cache contained the block, this was accounted for in its refcount. After this change, the refcount is only used to track external references. There is a boolean flag which indicates whether or not the block is contained in the cache. This diff also changes how LRU list is used. Previously, both hashtable and the LRU list contained all blocks. After this change, the LRU list contains blocks with the refcount==0, ie those which can be evicted from the cache. Note that this change still allows for cache to grow beyond its capacity. This happens when all blocks are pinned (ie refcount>0). This is consistent with the current behavior. The cache's insert function never fails. I spent lots of time trying to make table_reader and other places work with the insert which might failed. It turned out to be pretty hard. It might really destabilize some customers, so finally, I decided against doing this. table_cache_remove_scan_count_limit option will be unneeded after this change, but I will remove it in the following diff, if this one gets approved Test Plan: Ran tests, made sure they pass Reviewers: sdong, ljin Differential Revision: https://reviews.facebook.net/D25503
10 years ago
// cache is under capacity now since elements were released
ASSERT_EQ(n, cache->GetUsage());
// element 0 is evicted and the rest is there
// This is consistent with the LRU policy since the element 0
// was released first
for (int i = 0; i < static_cast<int>(n + 1); i++) {
std::string key = EncodeKey(i + 1);
Modifed the LRU cache eviction code so that it doesn't evict blocks which have exteranl references Summary: Currently, blocks which have more than one reference (ie referenced by something other than cache itself) are evicted from cache. This doesn't make much sense: - blocks are still in RAM, so the RAM usage reported by the cache is incorrect - if the same block is needed by another iterator, it will be loaded and decompressed again This diff changes the reference counting scheme a bit. Previously, if the cache contained the block, this was accounted for in its refcount. After this change, the refcount is only used to track external references. There is a boolean flag which indicates whether or not the block is contained in the cache. This diff also changes how LRU list is used. Previously, both hashtable and the LRU list contained all blocks. After this change, the LRU list contains blocks with the refcount==0, ie those which can be evicted from the cache. Note that this change still allows for cache to grow beyond its capacity. This happens when all blocks are pinned (ie refcount>0). This is consistent with the current behavior. The cache's insert function never fails. I spent lots of time trying to make table_reader and other places work with the insert which might failed. It turned out to be pretty hard. It might really destabilize some customers, so finally, I decided against doing this. table_cache_remove_scan_count_limit option will be unneeded after this change, but I will remove it in the following diff, if this one gets approved Test Plan: Ran tests, made sure they pass Reviewers: sdong, ljin Differential Revision: https://reviews.facebook.net/D25503
10 years ago
auto h = cache->Lookup(key);
if (h) {
ASSERT_NE(static_cast<size_t>(i), 0U);
Modifed the LRU cache eviction code so that it doesn't evict blocks which have exteranl references Summary: Currently, blocks which have more than one reference (ie referenced by something other than cache itself) are evicted from cache. This doesn't make much sense: - blocks are still in RAM, so the RAM usage reported by the cache is incorrect - if the same block is needed by another iterator, it will be loaded and decompressed again This diff changes the reference counting scheme a bit. Previously, if the cache contained the block, this was accounted for in its refcount. After this change, the refcount is only used to track external references. There is a boolean flag which indicates whether or not the block is contained in the cache. This diff also changes how LRU list is used. Previously, both hashtable and the LRU list contained all blocks. After this change, the LRU list contains blocks with the refcount==0, ie those which can be evicted from the cache. Note that this change still allows for cache to grow beyond its capacity. This happens when all blocks are pinned (ie refcount>0). This is consistent with the current behavior. The cache's insert function never fails. I spent lots of time trying to make table_reader and other places work with the insert which might failed. It turned out to be pretty hard. It might really destabilize some customers, so finally, I decided against doing this. table_cache_remove_scan_count_limit option will be unneeded after this change, but I will remove it in the following diff, if this one gets approved Test Plan: Ran tests, made sure they pass Reviewers: sdong, ljin Differential Revision: https://reviews.facebook.net/D25503
10 years ago
cache->Release(h);
} else {
ASSERT_EQ(static_cast<size_t>(i), 0U);
Modifed the LRU cache eviction code so that it doesn't evict blocks which have exteranl references Summary: Currently, blocks which have more than one reference (ie referenced by something other than cache itself) are evicted from cache. This doesn't make much sense: - blocks are still in RAM, so the RAM usage reported by the cache is incorrect - if the same block is needed by another iterator, it will be loaded and decompressed again This diff changes the reference counting scheme a bit. Previously, if the cache contained the block, this was accounted for in its refcount. After this change, the refcount is only used to track external references. There is a boolean flag which indicates whether or not the block is contained in the cache. This diff also changes how LRU list is used. Previously, both hashtable and the LRU list contained all blocks. After this change, the LRU list contains blocks with the refcount==0, ie those which can be evicted from the cache. Note that this change still allows for cache to grow beyond its capacity. This happens when all blocks are pinned (ie refcount>0). This is consistent with the current behavior. The cache's insert function never fails. I spent lots of time trying to make table_reader and other places work with the insert which might failed. It turned out to be pretty hard. It might really destabilize some customers, so finally, I decided against doing this. table_cache_remove_scan_count_limit option will be unneeded after this change, but I will remove it in the following diff, if this one gets approved Test Plan: Ran tests, made sure they pass Reviewers: sdong, ljin Differential Revision: https://reviews.facebook.net/D25503
10 years ago
}
}
}
namespace {
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
std::vector<std::pair<int, int>> legacy_callback_state;
void legacy_callback(void* value, size_t charge) {
legacy_callback_state.push_back(
{DecodeValue(value), static_cast<int>(charge)});
}
};
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
TEST_P(CacheTest, ApplyToAllCacheEntriesTest) {
std::vector<std::pair<int, int>> inserted;
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
legacy_callback_state.clear();
for (int i = 0; i < 10; ++i) {
Insert(i, i * 2, i + 1);
inserted.push_back({i * 2, i + 1});
}
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
cache_->ApplyToAllCacheEntries(legacy_callback, true);
std::sort(inserted.begin(), inserted.end());
std::sort(legacy_callback_state.begin(), legacy_callback_state.end());
ASSERT_EQ(inserted.size(), legacy_callback_state.size());
for (int i = 0; i < static_cast<int>(inserted.size()); ++i) {
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
EXPECT_EQ(inserted[i], legacy_callback_state[i]);
}
}
TEST_P(CacheTest, ApplyToAllEntriesTest) {
std::vector<std::string> callback_state;
const auto callback = [&](const Slice& key, void* value, size_t charge,
Cache::DeleterFn deleter) {
callback_state.push_back(std::to_string(DecodeKey(key)) + "," +
std::to_string(DecodeValue(value)) + "," +
std::to_string(charge));
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
assert(deleter == &CacheTest::Deleter);
};
std::vector<std::string> inserted;
callback_state.clear();
for (int i = 0; i < 10; ++i) {
Insert(i, i * 2, i + 1);
inserted.push_back(std::to_string(i) + "," + std::to_string(i * 2) + "," +
std::to_string(i + 1));
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
}
cache_->ApplyToAllEntries(callback, /*opts*/ {});
std::sort(inserted.begin(), inserted.end());
std::sort(callback_state.begin(), callback_state.end());
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
ASSERT_EQ(inserted.size(), callback_state.size());
for (int i = 0; i < static_cast<int>(inserted.size()); ++i) {
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
EXPECT_EQ(inserted[i], callback_state[i]);
}
}
TEST_P(CacheTest, ApplyToAllEntriesDuringResize) {
// This is a mini-stress test of ApplyToAllEntries, to ensure
// items in the cache that are neither added nor removed
// during ApplyToAllEntries are counted exactly once.
// Insert some entries that we expect to be seen exactly once
// during iteration.
constexpr int kSpecialCharge = 2;
constexpr int kNotSpecialCharge = 1;
constexpr int kSpecialCount = 100;
size_t expected_usage = 0;
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
for (int i = 0; i < kSpecialCount; ++i) {
Insert(i, i * 2, kSpecialCharge);
expected_usage += kSpecialCharge;
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
}
// For callback
int special_count = 0;
const auto callback = [&](const Slice&, void*, size_t charge,
Cache::DeleterFn) {
if (charge == static_cast<size_t>(kSpecialCharge)) {
++special_count;
}
};
// Start counting
std::thread apply_thread([&]() {
// Use small average_entries_per_lock to make the problem difficult
Cache::ApplyToAllEntriesOptions opts;
opts.average_entries_per_lock = 2;
cache_->ApplyToAllEntries(callback, opts);
});
// In parallel, add more entries, enough to cause resize but not enough
// to cause ejections. (Note: if any cache shard is over capacity, there
// will be ejections)
for (int i = kSpecialCount * 1; i < kSpecialCount * 5; ++i) {
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
Insert(i, i * 2, kNotSpecialCharge);
expected_usage += kNotSpecialCharge;
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
}
apply_thread.join();
// verify no evictions
ASSERT_EQ(cache_->GetUsage(), expected_usage);
// verify everything seen in ApplyToAllEntries
New Cache API for gathering statistics (#8225) Summary: Adds a new Cache::ApplyToAllEntries API that we expect to use (in follow-up PRs) for efficiently gathering block cache statistics. Notable features vs. old ApplyToAllCacheEntries: * Includes key and deleter (in addition to value and charge). We could have passed in a Handle but then more virtual function calls would be needed to get the "fields" of each entry. We expect to use the 'deleter' to identify the origin of entries, perhaps even more. * Heavily tuned to minimize latency impact on operating cache. It does this by iterating over small sections of each cache shard while cycling through the shards. * Supports tuning roughly how many entries to operate on for each lock acquire and release, to control the impact on the latency of other operations without excessive lock acquire & release. The right balance can depend on the cost of the callback. Good default seems to be around 256. * There should be no need to disable thread safety. (I would expect uncontended locks to be sufficiently fast.) I have enhanced cache_bench to validate this approach: * Reports a histogram of ns per operation, so we can look at the ditribution of times, not just throughput (average). * Can add a thread for simulated "gather stats" which calls ApplyToAllEntries at a specified interval. We also generate a histogram of time to run ApplyToAllEntries. To make the iteration over some entries of each shard work as cleanly as possible, even with resize between next set of entries, I have re-arranged which hash bits are used for sharding and which for indexing within a shard. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225 Test Plan: A couple of unit tests are added, but primary validation is manual, as the primary risk is to performance. The primary validation is using cache_bench to ensure that neither the minor hashing changes nor the simulated stats gathering significantly impact QPS or latency distribution. Note that adding op latency histogram seriously impacts the benchmark QPS, so for a fair baseline, we need the cache_bench changes (except remove simulated stat gathering to make it compile). In short, we don't see any reproducible difference in ops/sec or op latency unless we are gathering stats nearly continuously. Test uses 10GB block cache with 8KB values to be somewhat realistic in the number of items to iterate over. Baseline typical output: ``` Complete in 92.017 s; Rough parallel ops/sec = 869401 Thread ops/sec = 54662 Operation latency (ns): Count: 80000000 Average: 11223.9494 StdDev: 29.61 Min: 0 Median: 7759.3973 Max: 9620500 Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58 ------------------------------------------------------ [ 0, 1 ] 68 0.000% 0.000% ( 2900, 4400 ] 89 0.000% 0.000% ( 4400, 6600 ] 33630240 42.038% 42.038% ######## ( 6600, 9900 ] 18129842 22.662% 64.700% ##### ( 9900, 14000 ] 7877533 9.847% 74.547% ## ( 14000, 22000 ] 15193238 18.992% 93.539% #### ( 22000, 33000 ] 3037061 3.796% 97.335% # ( 33000, 50000 ] 1626316 2.033% 99.368% ( 50000, 75000 ] 421532 0.527% 99.895% ( 75000, 110000 ] 56910 0.071% 99.966% ( 110000, 170000 ] 16134 0.020% 99.986% ( 170000, 250000 ] 5166 0.006% 99.993% ( 250000, 380000 ] 3017 0.004% 99.996% ( 380000, 570000 ] 1337 0.002% 99.998% ( 570000, 860000 ] 805 0.001% 99.999% ( 860000, 1200000 ] 319 0.000% 100.000% ( 1200000, 1900000 ] 231 0.000% 100.000% ( 1900000, 2900000 ] 100 0.000% 100.000% ( 2900000, 4300000 ] 39 0.000% 100.000% ( 4300000, 6500000 ] 16 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ``` New, gather_stats=false. Median thread ops/sec of 5 runs: ``` Complete in 92.030 s; Rough parallel ops/sec = 869285 Thread ops/sec = 54458 Operation latency (ns): Count: 80000000 Average: 11298.1027 StdDev: 42.18 Min: 0 Median: 7722.0822 Max: 6398720 Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78 ------------------------------------------------------ [ 0, 1 ] 109 0.000% 0.000% ( 2900, 4400 ] 793 0.001% 0.001% ( 4400, 6600 ] 34054563 42.568% 42.569% ######### ( 6600, 9900 ] 17482646 21.853% 64.423% #### ( 9900, 14000 ] 7908180 9.885% 74.308% ## ( 14000, 22000 ] 15032072 18.790% 93.098% #### ( 22000, 33000 ] 3237834 4.047% 97.145% # ( 33000, 50000 ] 1736882 2.171% 99.316% ( 50000, 75000 ] 446851 0.559% 99.875% ( 75000, 110000 ] 68251 0.085% 99.960% ( 110000, 170000 ] 18592 0.023% 99.983% ( 170000, 250000 ] 7200 0.009% 99.992% ( 250000, 380000 ] 3334 0.004% 99.997% ( 380000, 570000 ] 1393 0.002% 99.998% ( 570000, 860000 ] 700 0.001% 99.999% ( 860000, 1200000 ] 293 0.000% 100.000% ( 1200000, 1900000 ] 196 0.000% 100.000% ( 1900000, 2900000 ] 69 0.000% 100.000% ( 2900000, 4300000 ] 32 0.000% 100.000% ( 4300000, 6500000 ] 10 0.000% 100.000% ``` New, gather_stats=true, 1 second delay between scans. Scans take about 1 second here so it's spending about 50% time scanning. Still the effect on ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs: ``` Complete in 91.890 s; Rough parallel ops/sec = 870608 Thread ops/sec = 54551 Operation latency (ns): Count: 80000000 Average: 11311.2629 StdDev: 45.28 Min: 0 Median: 7686.5458 Max: 10018340 Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86 ------------------------------------------------------ [ 0, 1 ] 71 0.000% 0.000% ( 2900, 4400 ] 291 0.000% 0.000% ( 4400, 6600 ] 34492060 43.115% 43.116% ######### ( 6600, 9900 ] 16727328 20.909% 64.025% #### ( 9900, 14000 ] 7845828 9.807% 73.832% ## ( 14000, 22000 ] 15510654 19.388% 93.220% #### ( 22000, 33000 ] 3216533 4.021% 97.241% # ( 33000, 50000 ] 1680859 2.101% 99.342% ( 50000, 75000 ] 439059 0.549% 99.891% ( 75000, 110000 ] 60540 0.076% 99.967% ( 110000, 170000 ] 14649 0.018% 99.985% ( 170000, 250000 ] 5242 0.007% 99.991% ( 250000, 380000 ] 3260 0.004% 99.995% ( 380000, 570000 ] 1599 0.002% 99.997% ( 570000, 860000 ] 1043 0.001% 99.999% ( 860000, 1200000 ] 471 0.001% 99.999% ( 1200000, 1900000 ] 275 0.000% 100.000% ( 1900000, 2900000 ] 143 0.000% 100.000% ( 2900000, 4300000 ] 60 0.000% 100.000% ( 4300000, 6500000 ] 27 0.000% 100.000% ( 6500000, 9800000 ] 7 0.000% 100.000% ( 9800000, 14000000 ] 1 0.000% 100.000% Gather stats latency (us): Count: 46 Average: 980387.5870 StdDev: 60911.18 Min: 879155 Median: 1033777.7778 Max: 1261431 Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00 ------------------------------------------------------ ( 860000, 1200000 ] 45 97.826% 97.826% #################### ( 1200000, 1900000 ] 1 2.174% 100.000% Most recent cache entry stats: Number of entries: 1295133 Total charge: 9.88 GB Average key size: 23.4982 Average charge: 8.00 KB Unique deleters: 3 ``` Reviewed By: mrambacher Differential Revision: D28295742 Pulled By: pdillinger fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
ASSERT_EQ(special_count, kSpecialCount);
}
TEST_P(CacheTest, DefaultShardBits) {
// test1: set the flag to false. Insert more keys than capacity. See if they
// all go through.
std::shared_ptr<Cache> cache = NewCache(16 * 1024L * 1024L);
ShardedCache* sc = dynamic_cast<ShardedCache*>(cache.get());
ASSERT_EQ(5, sc->GetNumShardBits());
cache = NewLRUCache(511 * 1024L, -1, true);
sc = dynamic_cast<ShardedCache*>(cache.get());
ASSERT_EQ(0, sc->GetNumShardBits());
cache = NewLRUCache(1024L * 1024L * 1024L, -1, true);
sc = dynamic_cast<ShardedCache*>(cache.get());
ASSERT_EQ(6, sc->GetNumShardBits());
}
Use deleters to label cache entries and collect stats (#8297) Summary: This change gathers and publishes statistics about the kinds of items in block cache. This is especially important for profiling relative usage of cache by index vs. filter vs. data blocks. It works by iterating over the cache during periodic stats dump (InternalStats, stats_dump_period_sec) or on demand when DB::Get(Map)Property(kBlockCacheEntryStats), except that for efficiency and sharing among column families, saved data from the last scan is used when the data is not considered too old. The new information can be seen in info LOG, for example: Block cache LRUCache@0x7fca62229330 capacity: 95.37 MB collections: 8 last_copies: 0 last_secs: 0.00178 secs_since: 0 Block cache entry stats(count,size,portion): DataBlock(7092,28.24 MB,29.6136%) FilterBlock(215,867.90 KB,0.888728%) FilterMetaBlock(2,5.31 KB,0.00544%) IndexBlock(217,180.11 KB,0.184432%) WriteBuffer(1,256.00 KB,0.262144%) Misc(1,0.00 KB,0%) And also through DB::GetProperty and GetMapProperty (here using ldb just for demonstration): $ ./ldb --db=/dev/shm/dbbench/ get_property rocksdb.block-cache-entry-stats rocksdb.block-cache-entry-stats.bytes.data-block: 0 rocksdb.block-cache-entry-stats.bytes.deprecated-filter-block: 0 rocksdb.block-cache-entry-stats.bytes.filter-block: 0 rocksdb.block-cache-entry-stats.bytes.filter-meta-block: 0 rocksdb.block-cache-entry-stats.bytes.index-block: 178992 rocksdb.block-cache-entry-stats.bytes.misc: 0 rocksdb.block-cache-entry-stats.bytes.other-block: 0 rocksdb.block-cache-entry-stats.bytes.write-buffer: 0 rocksdb.block-cache-entry-stats.capacity: 8388608 rocksdb.block-cache-entry-stats.count.data-block: 0 rocksdb.block-cache-entry-stats.count.deprecated-filter-block: 0 rocksdb.block-cache-entry-stats.count.filter-block: 0 rocksdb.block-cache-entry-stats.count.filter-meta-block: 0 rocksdb.block-cache-entry-stats.count.index-block: 215 rocksdb.block-cache-entry-stats.count.misc: 1 rocksdb.block-cache-entry-stats.count.other-block: 0 rocksdb.block-cache-entry-stats.count.write-buffer: 0 rocksdb.block-cache-entry-stats.id: LRUCache@0x7f3636661290 rocksdb.block-cache-entry-stats.percent.data-block: 0.000000 rocksdb.block-cache-entry-stats.percent.deprecated-filter-block: 0.000000 rocksdb.block-cache-entry-stats.percent.filter-block: 0.000000 rocksdb.block-cache-entry-stats.percent.filter-meta-block: 0.000000 rocksdb.block-cache-entry-stats.percent.index-block: 2.133751 rocksdb.block-cache-entry-stats.percent.misc: 0.000000 rocksdb.block-cache-entry-stats.percent.other-block: 0.000000 rocksdb.block-cache-entry-stats.percent.write-buffer: 0.000000 rocksdb.block-cache-entry-stats.secs_for_last_collection: 0.000052 rocksdb.block-cache-entry-stats.secs_since_last_collection: 0 Solution detail - We need some way to flag what kind of blocks each entry belongs to, preferably without changing the Cache API. One of the complications is that Cache is a general interface that could have other users that don't adhere to whichever convention we decide on for keys and values. Or we would pay for an extra field in the Handle that would only be used for this purpose. This change uses a back-door approach, the deleter, to indicate the "role" of a Cache entry (in addition to the value type, implicitly). This has the added benefit of ensuring proper code origin whenever we recognize a particular role for a cache entry; if the entry came from some other part of the code, it will use an unrecognized deleter, which we simply attribute to the "Misc" role. An internal API makes for simple instantiation and automatic registration of Cache deleters for a given value type and "role". Another internal API, CacheEntryStatsCollector, solves the problem of caching the results of a scan and sharing them, to ensure scans are neither excessive nor redundant so as not to harm Cache performance. Because code is added to BlocklikeTraits, it is pulled out of block_based_table_reader.cc into its own file. This is a reformulation of https://github.com/facebook/rocksdb/issues/8276, without the type checking option (could still be added), and with actual stat gathering. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8297 Test Plan: manual testing with db_bench, and a couple of basic unit tests Reviewed By: ltamasi Differential Revision: D28488721 Pulled By: pdillinger fbshipit-source-id: 472f524a9691b5afb107934be2d41d84f2b129fb
4 years ago
TEST_P(CacheTest, GetChargeAndDeleter) {
Insert(1, 2);
Cache::Handle* h1 = cache_->Lookup(EncodeKey(1));
ASSERT_EQ(2, DecodeValue(cache_->Value(h1)));
ASSERT_EQ(1, cache_->GetCharge(h1));
Use deleters to label cache entries and collect stats (#8297) Summary: This change gathers and publishes statistics about the kinds of items in block cache. This is especially important for profiling relative usage of cache by index vs. filter vs. data blocks. It works by iterating over the cache during periodic stats dump (InternalStats, stats_dump_period_sec) or on demand when DB::Get(Map)Property(kBlockCacheEntryStats), except that for efficiency and sharing among column families, saved data from the last scan is used when the data is not considered too old. The new information can be seen in info LOG, for example: Block cache LRUCache@0x7fca62229330 capacity: 95.37 MB collections: 8 last_copies: 0 last_secs: 0.00178 secs_since: 0 Block cache entry stats(count,size,portion): DataBlock(7092,28.24 MB,29.6136%) FilterBlock(215,867.90 KB,0.888728%) FilterMetaBlock(2,5.31 KB,0.00544%) IndexBlock(217,180.11 KB,0.184432%) WriteBuffer(1,256.00 KB,0.262144%) Misc(1,0.00 KB,0%) And also through DB::GetProperty and GetMapProperty (here using ldb just for demonstration): $ ./ldb --db=/dev/shm/dbbench/ get_property rocksdb.block-cache-entry-stats rocksdb.block-cache-entry-stats.bytes.data-block: 0 rocksdb.block-cache-entry-stats.bytes.deprecated-filter-block: 0 rocksdb.block-cache-entry-stats.bytes.filter-block: 0 rocksdb.block-cache-entry-stats.bytes.filter-meta-block: 0 rocksdb.block-cache-entry-stats.bytes.index-block: 178992 rocksdb.block-cache-entry-stats.bytes.misc: 0 rocksdb.block-cache-entry-stats.bytes.other-block: 0 rocksdb.block-cache-entry-stats.bytes.write-buffer: 0 rocksdb.block-cache-entry-stats.capacity: 8388608 rocksdb.block-cache-entry-stats.count.data-block: 0 rocksdb.block-cache-entry-stats.count.deprecated-filter-block: 0 rocksdb.block-cache-entry-stats.count.filter-block: 0 rocksdb.block-cache-entry-stats.count.filter-meta-block: 0 rocksdb.block-cache-entry-stats.count.index-block: 215 rocksdb.block-cache-entry-stats.count.misc: 1 rocksdb.block-cache-entry-stats.count.other-block: 0 rocksdb.block-cache-entry-stats.count.write-buffer: 0 rocksdb.block-cache-entry-stats.id: LRUCache@0x7f3636661290 rocksdb.block-cache-entry-stats.percent.data-block: 0.000000 rocksdb.block-cache-entry-stats.percent.deprecated-filter-block: 0.000000 rocksdb.block-cache-entry-stats.percent.filter-block: 0.000000 rocksdb.block-cache-entry-stats.percent.filter-meta-block: 0.000000 rocksdb.block-cache-entry-stats.percent.index-block: 2.133751 rocksdb.block-cache-entry-stats.percent.misc: 0.000000 rocksdb.block-cache-entry-stats.percent.other-block: 0.000000 rocksdb.block-cache-entry-stats.percent.write-buffer: 0.000000 rocksdb.block-cache-entry-stats.secs_for_last_collection: 0.000052 rocksdb.block-cache-entry-stats.secs_since_last_collection: 0 Solution detail - We need some way to flag what kind of blocks each entry belongs to, preferably without changing the Cache API. One of the complications is that Cache is a general interface that could have other users that don't adhere to whichever convention we decide on for keys and values. Or we would pay for an extra field in the Handle that would only be used for this purpose. This change uses a back-door approach, the deleter, to indicate the "role" of a Cache entry (in addition to the value type, implicitly). This has the added benefit of ensuring proper code origin whenever we recognize a particular role for a cache entry; if the entry came from some other part of the code, it will use an unrecognized deleter, which we simply attribute to the "Misc" role. An internal API makes for simple instantiation and automatic registration of Cache deleters for a given value type and "role". Another internal API, CacheEntryStatsCollector, solves the problem of caching the results of a scan and sharing them, to ensure scans are neither excessive nor redundant so as not to harm Cache performance. Because code is added to BlocklikeTraits, it is pulled out of block_based_table_reader.cc into its own file. This is a reformulation of https://github.com/facebook/rocksdb/issues/8276, without the type checking option (could still be added), and with actual stat gathering. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8297 Test Plan: manual testing with db_bench, and a couple of basic unit tests Reviewed By: ltamasi Differential Revision: D28488721 Pulled By: pdillinger fbshipit-source-id: 472f524a9691b5afb107934be2d41d84f2b129fb
4 years ago
ASSERT_EQ(&CacheTest::Deleter, cache_->GetDeleter(h1));
cache_->Release(h1);
}
#ifdef SUPPORT_CLOCK_CACHE
std::shared_ptr<Cache> (*new_clock_cache_func)(
size_t, int, bool, CacheMetadataChargePolicy) = NewClockCache;
INSTANTIATE_TEST_CASE_P(CacheTestInstance, CacheTest,
testing::Values(kLRU, kClock, kFast));
#else
INSTANTIATE_TEST_CASE_P(CacheTestInstance, CacheTest,
testing::Values(kLRU, kFast));
#endif // SUPPORT_CLOCK_CACHE
INSTANTIATE_TEST_CASE_P(CacheTestInstance, LRUCacheTest,
testing::Values(kLRU, kFast));
} // namespace ROCKSDB_NAMESPACE
int main(int argc, char** argv) {
ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
rocksdb: switch to gtest Summary: Our existing test notation is very similar to what is used in gtest. It makes it easy to adopt what is different. In this diff I modify existing [[ https://code.google.com/p/googletest/wiki/Primer#Test_Fixtures:_Using_the_Same_Data_Configuration_for_Multiple_Te | test fixture ]] classes to inherit from `testing::Test`. Also for unit tests that use fixture class, `TEST` is replaced with `TEST_F` as required in gtest. There are several custom `main` functions in our existing tests. To make this transition easier, I modify all `main` functions to fallow gtest notation. But eventually we can remove them and use implementation of `main` that gtest provides. ```lang=bash % cat ~/transform #!/bin/sh files=$(git ls-files '*test\.cc') for file in $files do if grep -q "rocksdb::test::RunAllTests()" $file then if grep -Eq '^class \w+Test {' $file then perl -pi -e 's/^(class \w+Test) {/${1}: public testing::Test {/g' $file perl -pi -e 's/^(TEST)/${1}_F/g' $file fi perl -pi -e 's/(int main.*\{)/${1}::testing::InitGoogleTest(&argc, argv);/g' $file perl -pi -e 's/rocksdb::test::RunAllTests/RUN_ALL_TESTS/g' $file fi done % sh ~/transform % make format ``` Second iteration of this diff contains only scripted changes. Third iteration contains manual changes to fix last errors and make it compilable. Test Plan: Build and notice no errors. ```lang=bash % USE_CLANG=1 make check -j55 ``` Tests are still testing. Reviewers: meyering, sdong, rven, igor Reviewed By: igor Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D35157
10 years ago
::testing::InitGoogleTest(&argc, argv);
return RUN_ALL_TESTS();
}