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/util/rate_limiter.cc

498 lines
18 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 "util/rate_limiter.h"
#include <algorithm>
#include "monitoring/statistics.h"
#include "port/port.h"
#include "rocksdb/convenience.h"
#include "rocksdb/system_clock.h"
#include "rocksdb/utilities/customizable_util.h"
#include "rocksdb/utilities/object_registry.h"
#include "rocksdb/utilities/options_type.h"
#include "test_util/sync_point.h"
#include "util/aligned_buffer.h"
#include "util/string_util.h"
namespace ROCKSDB_NAMESPACE {
size_t RateLimiter::RequestToken(size_t bytes, size_t alignment,
Env::IOPriority io_priority, Statistics* stats,
RateLimiter::OpType op_type) {
if (io_priority < Env::IO_TOTAL && IsRateLimited(op_type)) {
bytes = std::min(bytes, static_cast<size_t>(GetSingleBurstBytes()));
if (alignment > 0) {
// Here we may actually require more than burst and block
// as we can not write/read less than one page at a time on direct I/O
// thus we do not want to be strictly constrained by burst
bytes = std::max(alignment, TruncateToPageBoundary(alignment, bytes));
}
Request(bytes, io_priority, stats, op_type);
}
return bytes;
}
// Pending request
struct GenericRateLimiter::Req {
explicit Req(int64_t _bytes, port::Mutex* _mu)
fix rate limiter to avoid starvation Summary: The current implementation of rate limiter has the possibility to introduce resource starvation when change its limit. This diff aims to fix this problem by consuming request bytes partially. Test Plan: ``` ./rate_limiter_test [==========] Running 4 tests from 1 test case. [----------] Global test environment set-up. [----------] 4 tests from RateLimiterTest [ RUN ] RateLimiterTest.OverflowRate [ OK ] RateLimiterTest.OverflowRate (0 ms) [ RUN ] RateLimiterTest.StartStop [ OK ] RateLimiterTest.StartStop (0 ms) [ RUN ] RateLimiterTest.Rate request size [1 - 1023], limit 10 KB/sec, actual rate: 10.355712 KB/sec, elapsed 2.00 seconds request size [1 - 1023], limit 20 KB/sec, actual rate: 19.136564 KB/sec, elapsed 2.00 seconds request size [1 - 2047], limit 20 KB/sec, actual rate: 20.783976 KB/sec, elapsed 2.10 seconds request size [1 - 2047], limit 40 KB/sec, actual rate: 39.308144 KB/sec, elapsed 2.10 seconds request size [1 - 4095], limit 40 KB/sec, actual rate: 40.318349 KB/sec, elapsed 2.20 seconds request size [1 - 4095], limit 80 KB/sec, actual rate: 79.667396 KB/sec, elapsed 2.20 seconds request size [1 - 8191], limit 80 KB/sec, actual rate: 81.807158 KB/sec, elapsed 2.30 seconds request size [1 - 8191], limit 160 KB/sec, actual rate: 160.659761 KB/sec, elapsed 2.20 seconds request size [1 - 16383], limit 160 KB/sec, actual rate: 160.700990 KB/sec, elapsed 3.00 seconds request size [1 - 16383], limit 320 KB/sec, actual rate: 317.639481 KB/sec, elapsed 2.50 seconds [ OK ] RateLimiterTest.Rate (22618 ms) [ RUN ] RateLimiterTest.LimitChangeTest [COMPLETE] request size 10 KB, new limit 20KB/sec, refill period 1000 ms [COMPLETE] request size 10 KB, new limit 5KB/sec, refill period 1000 ms [COMPLETE] request size 20 KB, new limit 40KB/sec, refill period 1000 ms [COMPLETE] request size 20 KB, new limit 10KB/sec, refill period 1000 ms [COMPLETE] request size 40 KB, new limit 80KB/sec, refill period 1000 ms [COMPLETE] request size 40 KB, new limit 20KB/sec, refill period 1000 ms [COMPLETE] request size 80 KB, new limit 160KB/sec, refill period 1000 ms [COMPLETE] request size 80 KB, new limit 40KB/sec, refill period 1000 ms [COMPLETE] request size 160 KB, new limit 320KB/sec, refill period 1000 ms [COMPLETE] request size 160 KB, new limit 80KB/sec, refill period 1000 ms [ OK ] RateLimiterTest.LimitChangeTest (5002 ms) [----------] 4 tests from RateLimiterTest (27620 ms total) [----------] Global test environment tear-down [==========] 4 tests from 1 test case ran. (27621 ms total) [ PASSED ] 4 tests. ``` Reviewers: sdong, IslamAbdelRahman, yiwu, andrewkr Reviewed By: andrewkr Subscribers: andrewkr, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D60207
8 years ago
: request_bytes(_bytes), bytes(_bytes), cv(_mu), granted(false) {}
int64_t request_bytes;
int64_t bytes;
port::CondVar cv;
bool granted;
};
static std::unordered_map<std::string, OptionTypeInfo>
generic_rate_limiter_type_info = {
#ifndef ROCKSDB_LITE
{"rate_bytes_per_sec",
{offsetof(struct GenericRateLimiter::GenericRateLimiterOptions,
max_bytes_per_sec),
OptionType::kInt64T}},
{"refill_period_us",
{offsetof(struct GenericRateLimiter::GenericRateLimiterOptions,
refill_period_us),
OptionType::kInt64T}},
{"fairness",
{offsetof(struct GenericRateLimiter::GenericRateLimiterOptions,
fairness),
OptionType::kInt32T}},
{"auto_tuned",
{offsetof(struct GenericRateLimiter::GenericRateLimiterOptions,
auto_tuned),
OptionType::kBoolean}},
{"clock",
OptionTypeInfo::AsCustomSharedPtr<SystemClock>(
offsetof(struct GenericRateLimiter::GenericRateLimiterOptions,
clock),
OptionVerificationType::kByNameAllowFromNull,
OptionTypeFlags::kAllowNull)},
#endif // ROCKSDB_LITE
};
GenericRateLimiter::GenericRateLimiter(
int64_t rate_bytes_per_sec, int64_t refill_period_us, int32_t fairness,
RateLimiter::Mode mode, const std::shared_ptr<SystemClock>& clock,
bool auto_tuned)
: RateLimiter(mode),
options_(rate_bytes_per_sec, refill_period_us, fairness, clock,
auto_tuned),
stop_(false),
exit_cv_(&request_mutex_),
requests_to_wait_(0),
available_bytes_(0),
rnd_((uint32_t)time(nullptr)),
Simplify GenericRateLimiter algorithm (#8602) Summary: `GenericRateLimiter` slow path handles requests that cannot be satisfied immediately. Such requests enter a queue, and their thread stays in `Request()` until they are granted or the rate limiter is stopped. These threads are responsible for unblocking themselves. The work to do so is split into two main duties. (1) Waiting for the next refill time. (2) Refilling the bytes and granting requests. Prior to this PR, the slow path logic involved a leader election algorithm to pick one thread to perform (1) followed by (2). It elected the thread whose request was at the front of the highest priority non-empty queue since that request was most likely to be granted. This algorithm was efficient in terms of reducing intermediate wakeups, which is a thread waking up only to resume waiting after finding its request is not granted. However, the conceptual complexity of this algorithm was too high. It took me a long time to draw a timeline to understand how it works for just one edge case yet there were so many. This PR drops the leader election to reduce conceptual complexity. Now, the two duties can be performed by whichever thread acquires the lock first. The risk of this change is increasing the number of intermediate wakeups, however, we took steps to mitigate that. - `wait_until_refill_pending_` flag ensures only one thread performs (1). This\ prevents the thundering herd problem at the next refill time. The remaining\ threads wait on their condition variable with an unbounded duration -- thus we\ must remember to notify them to ensure forward progress. - (1) is typically done by a thread at the front of a queue. This is trivial\ when the queues are initially empty as the first choice that arrives must be\ the only entry in its queue. When queues are initially non-empty, we achieve\ this by having (2) notify a thread at the front of a queue (preferring higher\ priority) to perform the next duty. - We do not require any additional wakeup for (2). Typically it will just be\ done by the thread that finished (1). Combined, the second and third bullet points above suggest the refill/granting will typically be done by a request at the front of its queue. This is important because one wakeup is saved when a granted request happens to be in an already running thread. Note there are a few cases that still lead to intermediate wakeup, however. The first two are existing issues that also apply to the old algorithm, however, the third (including both subpoints) is new. - No request may be granted (only possible when rate limit dynamically\ decreases). - Requests from a different queue may be granted. - (2) may be run by a non-front request thread causing it to not be granted even\ if some requests in that same queue are granted. It can happen for a couple\ (unlikely) reasons. - A new request may sneak in and grab the lock at the refill time, before the\ thread finishing (1) can wake up and grab it. - A new request may sneak in and grab the lock and execute (1) before (2)'s\ chosen candidate can wake up and grab the lock. Then that non-front request\ thread performing (1) can carry over to perform (2). Pull Request resolved: https://github.com/facebook/rocksdb/pull/8602 Test Plan: - Use existing tests. The edge cases listed in the comment are all performance\ related; I could not really think of any related to correctness. The logic\ looks the same whether a thread wakes up/finishes its work early/on-time/late,\ or whether the thread is chosen vs. "steals" the work. - Verified write throughput and CPU overhead are basically the same with and\ without this change, even in a rate limiter heavy workload: Test command: ``` $ rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -benchmarks=fillrandom -num_multi_db=64 -num_low_pri_threads=64 -num_high_pri_threads=64 -write_buffer_size=262144 -target_file_size_base=262144 -max_bytes_for_level_base=1048576 -rate_limiter_bytes_per_sec=16777216 -key_size=24 -value_size=1000 -num=10000 -compression_type=none -rate_limiter_refill_period_us=1000 ``` Results before this PR: ``` fillrandom : 108.463 micros/op 9219 ops/sec; 9.0 MB/s 7.40user 8.84system 1:26.20elapsed 18%CPU (0avgtext+0avgdata 256140maxresident)k ``` Results after this PR: ``` fillrandom : 108.108 micros/op 9250 ops/sec; 9.0 MB/s 7.45user 8.23system 1:26.68elapsed 18%CPU (0avgtext+0avgdata 255688maxresident)k ``` Reviewed By: hx235 Differential Revision: D30048013 Pulled By: ajkr fbshipit-source-id: 6741bba9d9dfbccab359806d725105817fef818b
3 years ago
wait_until_refill_pending_(false),
num_drains_(0) {
RegisterOptions(&options_, &generic_rate_limiter_type_info);
Implement superior user & mid IO priority level in GenericRateLimiter (#8595) Summary: Context: An extra IO_USER priority in rate limiter allows users to optionally charge WAL writes / SST reads to rate limiter at this priority level, which then has higher priority than IO_HIGH and IO_LOW. With an extra IO_USER priority, it allows users to better specify the relative urgency/importance among different requests in rate limiter. As a consequence, IO resource management can better prioritize and limit resource based on user's need. The IO_USER is implemented as superior priority in GenericRateLimiter, in the sense that its request queue will always be iterated first without being constrained to fairness. The reason is that the notion of fairness is only meaningful in helping lower priorities in background IO (i.e, IO_HIGH/MID/LOW) to gain some fair chance to run so that it does not block foreground IO (i.e, the ones that are charged at the level of IO_USER). As we can see, the ultimate goal here is to not blocking foreground IO at IO_USER level, which justifies the superiority of IO_USER. Similar benefits exist for IO_MID priority. - Rewrote the logic of deciding the order of iterating request queues of high/low priorities to include the extra user/mid priority w/o affecting the existing behavior (see PR's [comment](https://github.com/facebook/rocksdb/pull/8595/files#r678749331)) - Included the request queue of user-pri/mid-pri in the code path of next-leader-candidate signaling and GenericRateLimiter's destructor - Included the extra user/mid-pri in bookkeeping data structures: total_bytes_through_ and total_requests_ - Re-written the previous impl of explicitly iterating priorities with a loop from Env::IO_LOW to Env::IO_TOTAL Pull Request resolved: https://github.com/facebook/rocksdb/pull/8595 Test Plan: - passed existing rate_limiter_test.cc - passed added unit tests in rate_limiter_test.cc - run performance test to verify performance with only high/low requests is not affected by this change - Set-up command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom --duration=5 --compression_type=none --num=100000000 --disable_auto_compactions=true --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1))` - Test command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=overwrite --use_existing_db=true --disable_wal=true --duration=30 --compression_type=none --num=100000000 --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1)) --statistics=true --rate_limiter_bytes_per_sec=1048576 --rate_limiter_refill_period_us=1000 --threads=32 |& grep -E '(flush|compact)\.write\.bytes'` - Before (on branch upstream/master): `rocksdb.compact.write.bytes COUNT : 4014162` `rocksdb.flush.write.bytes COUNT : 26715832` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.66 - After (on branch rate_limiter_user_pri): `rocksdb.compact.write.bytes COUNT : 3807822` `rocksdb.flush.write.bytes COUNT : 26098659` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.85 Reviewed By: ajkr Differential Revision: D30577783 Pulled By: hx235 fbshipit-source-id: 0881f2705ffd13ecd331256bde7e8ec874a353f4
3 years ago
for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
total_requests_[i] = 0;
total_bytes_through_[i] = 0;
}
Initialize();
}
void GenericRateLimiter::Initialize() {
if (options_.clock == nullptr) {
options_.clock = SystemClock::Default();
}
options_.fairness = std::min(options_.fairness, 100);
next_refill_us_ = NowMicrosMonotonic();
tuned_time_ = std::chrono::microseconds(NowMicrosMonotonic());
if (options_.auto_tuned) {
rate_bytes_per_sec_ = options_.max_bytes_per_sec / 2;
} else {
rate_bytes_per_sec_ = options_.max_bytes_per_sec;
}
refill_bytes_per_period_ = CalculateRefillBytesPerPeriod(rate_bytes_per_sec_);
}
GenericRateLimiter::~GenericRateLimiter() {
MutexLock g(&request_mutex_);
stop_ = true;
Implement superior user & mid IO priority level in GenericRateLimiter (#8595) Summary: Context: An extra IO_USER priority in rate limiter allows users to optionally charge WAL writes / SST reads to rate limiter at this priority level, which then has higher priority than IO_HIGH and IO_LOW. With an extra IO_USER priority, it allows users to better specify the relative urgency/importance among different requests in rate limiter. As a consequence, IO resource management can better prioritize and limit resource based on user's need. The IO_USER is implemented as superior priority in GenericRateLimiter, in the sense that its request queue will always be iterated first without being constrained to fairness. The reason is that the notion of fairness is only meaningful in helping lower priorities in background IO (i.e, IO_HIGH/MID/LOW) to gain some fair chance to run so that it does not block foreground IO (i.e, the ones that are charged at the level of IO_USER). As we can see, the ultimate goal here is to not blocking foreground IO at IO_USER level, which justifies the superiority of IO_USER. Similar benefits exist for IO_MID priority. - Rewrote the logic of deciding the order of iterating request queues of high/low priorities to include the extra user/mid priority w/o affecting the existing behavior (see PR's [comment](https://github.com/facebook/rocksdb/pull/8595/files#r678749331)) - Included the request queue of user-pri/mid-pri in the code path of next-leader-candidate signaling and GenericRateLimiter's destructor - Included the extra user/mid-pri in bookkeeping data structures: total_bytes_through_ and total_requests_ - Re-written the previous impl of explicitly iterating priorities with a loop from Env::IO_LOW to Env::IO_TOTAL Pull Request resolved: https://github.com/facebook/rocksdb/pull/8595 Test Plan: - passed existing rate_limiter_test.cc - passed added unit tests in rate_limiter_test.cc - run performance test to verify performance with only high/low requests is not affected by this change - Set-up command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom --duration=5 --compression_type=none --num=100000000 --disable_auto_compactions=true --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1))` - Test command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=overwrite --use_existing_db=true --disable_wal=true --duration=30 --compression_type=none --num=100000000 --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1)) --statistics=true --rate_limiter_bytes_per_sec=1048576 --rate_limiter_refill_period_us=1000 --threads=32 |& grep -E '(flush|compact)\.write\.bytes'` - Before (on branch upstream/master): `rocksdb.compact.write.bytes COUNT : 4014162` `rocksdb.flush.write.bytes COUNT : 26715832` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.66 - After (on branch rate_limiter_user_pri): `rocksdb.compact.write.bytes COUNT : 3807822` `rocksdb.flush.write.bytes COUNT : 26098659` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.85 Reviewed By: ajkr Differential Revision: D30577783 Pulled By: hx235 fbshipit-source-id: 0881f2705ffd13ecd331256bde7e8ec874a353f4
3 years ago
std::deque<Req*>::size_type queues_size_sum = 0;
for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
queues_size_sum += queue_[i].size();
}
Implement superior user & mid IO priority level in GenericRateLimiter (#8595) Summary: Context: An extra IO_USER priority in rate limiter allows users to optionally charge WAL writes / SST reads to rate limiter at this priority level, which then has higher priority than IO_HIGH and IO_LOW. With an extra IO_USER priority, it allows users to better specify the relative urgency/importance among different requests in rate limiter. As a consequence, IO resource management can better prioritize and limit resource based on user's need. The IO_USER is implemented as superior priority in GenericRateLimiter, in the sense that its request queue will always be iterated first without being constrained to fairness. The reason is that the notion of fairness is only meaningful in helping lower priorities in background IO (i.e, IO_HIGH/MID/LOW) to gain some fair chance to run so that it does not block foreground IO (i.e, the ones that are charged at the level of IO_USER). As we can see, the ultimate goal here is to not blocking foreground IO at IO_USER level, which justifies the superiority of IO_USER. Similar benefits exist for IO_MID priority. - Rewrote the logic of deciding the order of iterating request queues of high/low priorities to include the extra user/mid priority w/o affecting the existing behavior (see PR's [comment](https://github.com/facebook/rocksdb/pull/8595/files#r678749331)) - Included the request queue of user-pri/mid-pri in the code path of next-leader-candidate signaling and GenericRateLimiter's destructor - Included the extra user/mid-pri in bookkeeping data structures: total_bytes_through_ and total_requests_ - Re-written the previous impl of explicitly iterating priorities with a loop from Env::IO_LOW to Env::IO_TOTAL Pull Request resolved: https://github.com/facebook/rocksdb/pull/8595 Test Plan: - passed existing rate_limiter_test.cc - passed added unit tests in rate_limiter_test.cc - run performance test to verify performance with only high/low requests is not affected by this change - Set-up command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom --duration=5 --compression_type=none --num=100000000 --disable_auto_compactions=true --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1))` - Test command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=overwrite --use_existing_db=true --disable_wal=true --duration=30 --compression_type=none --num=100000000 --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1)) --statistics=true --rate_limiter_bytes_per_sec=1048576 --rate_limiter_refill_period_us=1000 --threads=32 |& grep -E '(flush|compact)\.write\.bytes'` - Before (on branch upstream/master): `rocksdb.compact.write.bytes COUNT : 4014162` `rocksdb.flush.write.bytes COUNT : 26715832` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.66 - After (on branch rate_limiter_user_pri): `rocksdb.compact.write.bytes COUNT : 3807822` `rocksdb.flush.write.bytes COUNT : 26098659` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.85 Reviewed By: ajkr Differential Revision: D30577783 Pulled By: hx235 fbshipit-source-id: 0881f2705ffd13ecd331256bde7e8ec874a353f4
3 years ago
requests_to_wait_ = static_cast<int32_t>(queues_size_sum);
for (int i = Env::IO_TOTAL - 1; i >= Env::IO_LOW; --i) {
std::deque<Req*> queue = queue_[i];
for (auto& r : queue) {
r->cv.Signal();
}
}
Implement superior user & mid IO priority level in GenericRateLimiter (#8595) Summary: Context: An extra IO_USER priority in rate limiter allows users to optionally charge WAL writes / SST reads to rate limiter at this priority level, which then has higher priority than IO_HIGH and IO_LOW. With an extra IO_USER priority, it allows users to better specify the relative urgency/importance among different requests in rate limiter. As a consequence, IO resource management can better prioritize and limit resource based on user's need. The IO_USER is implemented as superior priority in GenericRateLimiter, in the sense that its request queue will always be iterated first without being constrained to fairness. The reason is that the notion of fairness is only meaningful in helping lower priorities in background IO (i.e, IO_HIGH/MID/LOW) to gain some fair chance to run so that it does not block foreground IO (i.e, the ones that are charged at the level of IO_USER). As we can see, the ultimate goal here is to not blocking foreground IO at IO_USER level, which justifies the superiority of IO_USER. Similar benefits exist for IO_MID priority. - Rewrote the logic of deciding the order of iterating request queues of high/low priorities to include the extra user/mid priority w/o affecting the existing behavior (see PR's [comment](https://github.com/facebook/rocksdb/pull/8595/files#r678749331)) - Included the request queue of user-pri/mid-pri in the code path of next-leader-candidate signaling and GenericRateLimiter's destructor - Included the extra user/mid-pri in bookkeeping data structures: total_bytes_through_ and total_requests_ - Re-written the previous impl of explicitly iterating priorities with a loop from Env::IO_LOW to Env::IO_TOTAL Pull Request resolved: https://github.com/facebook/rocksdb/pull/8595 Test Plan: - passed existing rate_limiter_test.cc - passed added unit tests in rate_limiter_test.cc - run performance test to verify performance with only high/low requests is not affected by this change - Set-up command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom --duration=5 --compression_type=none --num=100000000 --disable_auto_compactions=true --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1))` - Test command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=overwrite --use_existing_db=true --disable_wal=true --duration=30 --compression_type=none --num=100000000 --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1)) --statistics=true --rate_limiter_bytes_per_sec=1048576 --rate_limiter_refill_period_us=1000 --threads=32 |& grep -E '(flush|compact)\.write\.bytes'` - Before (on branch upstream/master): `rocksdb.compact.write.bytes COUNT : 4014162` `rocksdb.flush.write.bytes COUNT : 26715832` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.66 - After (on branch rate_limiter_user_pri): `rocksdb.compact.write.bytes COUNT : 3807822` `rocksdb.flush.write.bytes COUNT : 26098659` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.85 Reviewed By: ajkr Differential Revision: D30577783 Pulled By: hx235 fbshipit-source-id: 0881f2705ffd13ecd331256bde7e8ec874a353f4
3 years ago
while (requests_to_wait_ > 0) {
exit_cv_.Wait();
}
}
Status GenericRateLimiter::PrepareOptions(const ConfigOptions& options) {
if (options_.fairness <= 0) {
return Status::InvalidArgument("Fairness must be > 0");
} else if (options_.max_bytes_per_sec <= 0) {
return Status::InvalidArgument("max_bytes_per_sec must be > 0");
} else if (options_.refill_period_us <= 0) {
return Status::InvalidArgument("Refill_period_us must be > 0");
}
Initialize();
return RateLimiter::PrepareOptions(options);
}
// This API allows user to dynamically change rate limiter's bytes per second.
void GenericRateLimiter::SetBytesPerSecond(int64_t bytes_per_second) {
assert(bytes_per_second > 0);
rate_bytes_per_sec_ = bytes_per_second;
refill_bytes_per_period_.store(
CalculateRefillBytesPerPeriod(bytes_per_second),
std::memory_order_relaxed);
}
void GenericRateLimiter::Request(int64_t bytes, const Env::IOPriority pri,
Statistics* stats) {
assert(bytes <= refill_bytes_per_period_.load(std::memory_order_relaxed));
bytes = std::max(static_cast<int64_t>(0), bytes);
fix rate limiter to avoid starvation Summary: The current implementation of rate limiter has the possibility to introduce resource starvation when change its limit. This diff aims to fix this problem by consuming request bytes partially. Test Plan: ``` ./rate_limiter_test [==========] Running 4 tests from 1 test case. [----------] Global test environment set-up. [----------] 4 tests from RateLimiterTest [ RUN ] RateLimiterTest.OverflowRate [ OK ] RateLimiterTest.OverflowRate (0 ms) [ RUN ] RateLimiterTest.StartStop [ OK ] RateLimiterTest.StartStop (0 ms) [ RUN ] RateLimiterTest.Rate request size [1 - 1023], limit 10 KB/sec, actual rate: 10.355712 KB/sec, elapsed 2.00 seconds request size [1 - 1023], limit 20 KB/sec, actual rate: 19.136564 KB/sec, elapsed 2.00 seconds request size [1 - 2047], limit 20 KB/sec, actual rate: 20.783976 KB/sec, elapsed 2.10 seconds request size [1 - 2047], limit 40 KB/sec, actual rate: 39.308144 KB/sec, elapsed 2.10 seconds request size [1 - 4095], limit 40 KB/sec, actual rate: 40.318349 KB/sec, elapsed 2.20 seconds request size [1 - 4095], limit 80 KB/sec, actual rate: 79.667396 KB/sec, elapsed 2.20 seconds request size [1 - 8191], limit 80 KB/sec, actual rate: 81.807158 KB/sec, elapsed 2.30 seconds request size [1 - 8191], limit 160 KB/sec, actual rate: 160.659761 KB/sec, elapsed 2.20 seconds request size [1 - 16383], limit 160 KB/sec, actual rate: 160.700990 KB/sec, elapsed 3.00 seconds request size [1 - 16383], limit 320 KB/sec, actual rate: 317.639481 KB/sec, elapsed 2.50 seconds [ OK ] RateLimiterTest.Rate (22618 ms) [ RUN ] RateLimiterTest.LimitChangeTest [COMPLETE] request size 10 KB, new limit 20KB/sec, refill period 1000 ms [COMPLETE] request size 10 KB, new limit 5KB/sec, refill period 1000 ms [COMPLETE] request size 20 KB, new limit 40KB/sec, refill period 1000 ms [COMPLETE] request size 20 KB, new limit 10KB/sec, refill period 1000 ms [COMPLETE] request size 40 KB, new limit 80KB/sec, refill period 1000 ms [COMPLETE] request size 40 KB, new limit 20KB/sec, refill period 1000 ms [COMPLETE] request size 80 KB, new limit 160KB/sec, refill period 1000 ms [COMPLETE] request size 80 KB, new limit 40KB/sec, refill period 1000 ms [COMPLETE] request size 160 KB, new limit 320KB/sec, refill period 1000 ms [COMPLETE] request size 160 KB, new limit 80KB/sec, refill period 1000 ms [ OK ] RateLimiterTest.LimitChangeTest (5002 ms) [----------] 4 tests from RateLimiterTest (27620 ms total) [----------] Global test environment tear-down [==========] 4 tests from 1 test case ran. (27621 ms total) [ PASSED ] 4 tests. ``` Reviewers: sdong, IslamAbdelRahman, yiwu, andrewkr Reviewed By: andrewkr Subscribers: andrewkr, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D60207
8 years ago
TEST_SYNC_POINT("GenericRateLimiter::Request");
TEST_SYNC_POINT_CALLBACK("GenericRateLimiter::Request:1",
&rate_bytes_per_sec_);
MutexLock g(&request_mutex_);
if (options_.auto_tuned) {
static const int kRefillsPerTune = 100;
std::chrono::microseconds now(NowMicrosMonotonic());
if (now - tuned_time_ >= kRefillsPerTune * std::chrono::microseconds(
options_.refill_period_us)) {
Status s = Tune();
s.PermitUncheckedError(); //**TODO: What to do on error?
}
}
if (stop_) {
// It is now in the clean-up of ~GenericRateLimiter().
// Therefore any new incoming request will exit from here
// and not get satiesfied.
return;
}
++total_requests_[pri];
if (available_bytes_ >= bytes) {
// Refill thread assigns quota and notifies requests waiting on
// the queue under mutex. So if we get here, that means nobody
// is waiting?
available_bytes_ -= bytes;
total_bytes_through_[pri] += bytes;
return;
}
// Request cannot be satisfied at this moment, enqueue
Req r(bytes, &request_mutex_);
queue_[pri].push_back(&r);
TEST_SYNC_POINT_CALLBACK("GenericRateLimiter::Request:PostEnqueueRequest",
&request_mutex_);
Simplify GenericRateLimiter algorithm (#8602) Summary: `GenericRateLimiter` slow path handles requests that cannot be satisfied immediately. Such requests enter a queue, and their thread stays in `Request()` until they are granted or the rate limiter is stopped. These threads are responsible for unblocking themselves. The work to do so is split into two main duties. (1) Waiting for the next refill time. (2) Refilling the bytes and granting requests. Prior to this PR, the slow path logic involved a leader election algorithm to pick one thread to perform (1) followed by (2). It elected the thread whose request was at the front of the highest priority non-empty queue since that request was most likely to be granted. This algorithm was efficient in terms of reducing intermediate wakeups, which is a thread waking up only to resume waiting after finding its request is not granted. However, the conceptual complexity of this algorithm was too high. It took me a long time to draw a timeline to understand how it works for just one edge case yet there were so many. This PR drops the leader election to reduce conceptual complexity. Now, the two duties can be performed by whichever thread acquires the lock first. The risk of this change is increasing the number of intermediate wakeups, however, we took steps to mitigate that. - `wait_until_refill_pending_` flag ensures only one thread performs (1). This\ prevents the thundering herd problem at the next refill time. The remaining\ threads wait on their condition variable with an unbounded duration -- thus we\ must remember to notify them to ensure forward progress. - (1) is typically done by a thread at the front of a queue. This is trivial\ when the queues are initially empty as the first choice that arrives must be\ the only entry in its queue. When queues are initially non-empty, we achieve\ this by having (2) notify a thread at the front of a queue (preferring higher\ priority) to perform the next duty. - We do not require any additional wakeup for (2). Typically it will just be\ done by the thread that finished (1). Combined, the second and third bullet points above suggest the refill/granting will typically be done by a request at the front of its queue. This is important because one wakeup is saved when a granted request happens to be in an already running thread. Note there are a few cases that still lead to intermediate wakeup, however. The first two are existing issues that also apply to the old algorithm, however, the third (including both subpoints) is new. - No request may be granted (only possible when rate limit dynamically\ decreases). - Requests from a different queue may be granted. - (2) may be run by a non-front request thread causing it to not be granted even\ if some requests in that same queue are granted. It can happen for a couple\ (unlikely) reasons. - A new request may sneak in and grab the lock at the refill time, before the\ thread finishing (1) can wake up and grab it. - A new request may sneak in and grab the lock and execute (1) before (2)'s\ chosen candidate can wake up and grab the lock. Then that non-front request\ thread performing (1) can carry over to perform (2). Pull Request resolved: https://github.com/facebook/rocksdb/pull/8602 Test Plan: - Use existing tests. The edge cases listed in the comment are all performance\ related; I could not really think of any related to correctness. The logic\ looks the same whether a thread wakes up/finishes its work early/on-time/late,\ or whether the thread is chosen vs. "steals" the work. - Verified write throughput and CPU overhead are basically the same with and\ without this change, even in a rate limiter heavy workload: Test command: ``` $ rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -benchmarks=fillrandom -num_multi_db=64 -num_low_pri_threads=64 -num_high_pri_threads=64 -write_buffer_size=262144 -target_file_size_base=262144 -max_bytes_for_level_base=1048576 -rate_limiter_bytes_per_sec=16777216 -key_size=24 -value_size=1000 -num=10000 -compression_type=none -rate_limiter_refill_period_us=1000 ``` Results before this PR: ``` fillrandom : 108.463 micros/op 9219 ops/sec; 9.0 MB/s 7.40user 8.84system 1:26.20elapsed 18%CPU (0avgtext+0avgdata 256140maxresident)k ``` Results after this PR: ``` fillrandom : 108.108 micros/op 9250 ops/sec; 9.0 MB/s 7.45user 8.23system 1:26.68elapsed 18%CPU (0avgtext+0avgdata 255688maxresident)k ``` Reviewed By: hx235 Differential Revision: D30048013 Pulled By: ajkr fbshipit-source-id: 6741bba9d9dfbccab359806d725105817fef818b
3 years ago
// A thread representing a queued request coordinates with other such threads.
// There are two main duties.
//
// (1) Waiting for the next refill time.
// (2) Refilling the bytes and granting requests.
do {
Simplify GenericRateLimiter algorithm (#8602) Summary: `GenericRateLimiter` slow path handles requests that cannot be satisfied immediately. Such requests enter a queue, and their thread stays in `Request()` until they are granted or the rate limiter is stopped. These threads are responsible for unblocking themselves. The work to do so is split into two main duties. (1) Waiting for the next refill time. (2) Refilling the bytes and granting requests. Prior to this PR, the slow path logic involved a leader election algorithm to pick one thread to perform (1) followed by (2). It elected the thread whose request was at the front of the highest priority non-empty queue since that request was most likely to be granted. This algorithm was efficient in terms of reducing intermediate wakeups, which is a thread waking up only to resume waiting after finding its request is not granted. However, the conceptual complexity of this algorithm was too high. It took me a long time to draw a timeline to understand how it works for just one edge case yet there were so many. This PR drops the leader election to reduce conceptual complexity. Now, the two duties can be performed by whichever thread acquires the lock first. The risk of this change is increasing the number of intermediate wakeups, however, we took steps to mitigate that. - `wait_until_refill_pending_` flag ensures only one thread performs (1). This\ prevents the thundering herd problem at the next refill time. The remaining\ threads wait on their condition variable with an unbounded duration -- thus we\ must remember to notify them to ensure forward progress. - (1) is typically done by a thread at the front of a queue. This is trivial\ when the queues are initially empty as the first choice that arrives must be\ the only entry in its queue. When queues are initially non-empty, we achieve\ this by having (2) notify a thread at the front of a queue (preferring higher\ priority) to perform the next duty. - We do not require any additional wakeup for (2). Typically it will just be\ done by the thread that finished (1). Combined, the second and third bullet points above suggest the refill/granting will typically be done by a request at the front of its queue. This is important because one wakeup is saved when a granted request happens to be in an already running thread. Note there are a few cases that still lead to intermediate wakeup, however. The first two are existing issues that also apply to the old algorithm, however, the third (including both subpoints) is new. - No request may be granted (only possible when rate limit dynamically\ decreases). - Requests from a different queue may be granted. - (2) may be run by a non-front request thread causing it to not be granted even\ if some requests in that same queue are granted. It can happen for a couple\ (unlikely) reasons. - A new request may sneak in and grab the lock at the refill time, before the\ thread finishing (1) can wake up and grab it. - A new request may sneak in and grab the lock and execute (1) before (2)'s\ chosen candidate can wake up and grab the lock. Then that non-front request\ thread performing (1) can carry over to perform (2). Pull Request resolved: https://github.com/facebook/rocksdb/pull/8602 Test Plan: - Use existing tests. The edge cases listed in the comment are all performance\ related; I could not really think of any related to correctness. The logic\ looks the same whether a thread wakes up/finishes its work early/on-time/late,\ or whether the thread is chosen vs. "steals" the work. - Verified write throughput and CPU overhead are basically the same with and\ without this change, even in a rate limiter heavy workload: Test command: ``` $ rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -benchmarks=fillrandom -num_multi_db=64 -num_low_pri_threads=64 -num_high_pri_threads=64 -write_buffer_size=262144 -target_file_size_base=262144 -max_bytes_for_level_base=1048576 -rate_limiter_bytes_per_sec=16777216 -key_size=24 -value_size=1000 -num=10000 -compression_type=none -rate_limiter_refill_period_us=1000 ``` Results before this PR: ``` fillrandom : 108.463 micros/op 9219 ops/sec; 9.0 MB/s 7.40user 8.84system 1:26.20elapsed 18%CPU (0avgtext+0avgdata 256140maxresident)k ``` Results after this PR: ``` fillrandom : 108.108 micros/op 9250 ops/sec; 9.0 MB/s 7.45user 8.23system 1:26.68elapsed 18%CPU (0avgtext+0avgdata 255688maxresident)k ``` Reviewed By: hx235 Differential Revision: D30048013 Pulled By: ajkr fbshipit-source-id: 6741bba9d9dfbccab359806d725105817fef818b
3 years ago
int64_t time_until_refill_us = next_refill_us_ - NowMicrosMonotonic();
if (time_until_refill_us > 0) {
if (wait_until_refill_pending_) {
// Somebody is performing (1). Trust we'll be woken up when our request
// is granted or we are needed for future duties.
r.cv.Wait();
} else {
Simplify GenericRateLimiter algorithm (#8602) Summary: `GenericRateLimiter` slow path handles requests that cannot be satisfied immediately. Such requests enter a queue, and their thread stays in `Request()` until they are granted or the rate limiter is stopped. These threads are responsible for unblocking themselves. The work to do so is split into two main duties. (1) Waiting for the next refill time. (2) Refilling the bytes and granting requests. Prior to this PR, the slow path logic involved a leader election algorithm to pick one thread to perform (1) followed by (2). It elected the thread whose request was at the front of the highest priority non-empty queue since that request was most likely to be granted. This algorithm was efficient in terms of reducing intermediate wakeups, which is a thread waking up only to resume waiting after finding its request is not granted. However, the conceptual complexity of this algorithm was too high. It took me a long time to draw a timeline to understand how it works for just one edge case yet there were so many. This PR drops the leader election to reduce conceptual complexity. Now, the two duties can be performed by whichever thread acquires the lock first. The risk of this change is increasing the number of intermediate wakeups, however, we took steps to mitigate that. - `wait_until_refill_pending_` flag ensures only one thread performs (1). This\ prevents the thundering herd problem at the next refill time. The remaining\ threads wait on their condition variable with an unbounded duration -- thus we\ must remember to notify them to ensure forward progress. - (1) is typically done by a thread at the front of a queue. This is trivial\ when the queues are initially empty as the first choice that arrives must be\ the only entry in its queue. When queues are initially non-empty, we achieve\ this by having (2) notify a thread at the front of a queue (preferring higher\ priority) to perform the next duty. - We do not require any additional wakeup for (2). Typically it will just be\ done by the thread that finished (1). Combined, the second and third bullet points above suggest the refill/granting will typically be done by a request at the front of its queue. This is important because one wakeup is saved when a granted request happens to be in an already running thread. Note there are a few cases that still lead to intermediate wakeup, however. The first two are existing issues that also apply to the old algorithm, however, the third (including both subpoints) is new. - No request may be granted (only possible when rate limit dynamically\ decreases). - Requests from a different queue may be granted. - (2) may be run by a non-front request thread causing it to not be granted even\ if some requests in that same queue are granted. It can happen for a couple\ (unlikely) reasons. - A new request may sneak in and grab the lock at the refill time, before the\ thread finishing (1) can wake up and grab it. - A new request may sneak in and grab the lock and execute (1) before (2)'s\ chosen candidate can wake up and grab the lock. Then that non-front request\ thread performing (1) can carry over to perform (2). Pull Request resolved: https://github.com/facebook/rocksdb/pull/8602 Test Plan: - Use existing tests. The edge cases listed in the comment are all performance\ related; I could not really think of any related to correctness. The logic\ looks the same whether a thread wakes up/finishes its work early/on-time/late,\ or whether the thread is chosen vs. "steals" the work. - Verified write throughput and CPU overhead are basically the same with and\ without this change, even in a rate limiter heavy workload: Test command: ``` $ rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -benchmarks=fillrandom -num_multi_db=64 -num_low_pri_threads=64 -num_high_pri_threads=64 -write_buffer_size=262144 -target_file_size_base=262144 -max_bytes_for_level_base=1048576 -rate_limiter_bytes_per_sec=16777216 -key_size=24 -value_size=1000 -num=10000 -compression_type=none -rate_limiter_refill_period_us=1000 ``` Results before this PR: ``` fillrandom : 108.463 micros/op 9219 ops/sec; 9.0 MB/s 7.40user 8.84system 1:26.20elapsed 18%CPU (0avgtext+0avgdata 256140maxresident)k ``` Results after this PR: ``` fillrandom : 108.108 micros/op 9250 ops/sec; 9.0 MB/s 7.45user 8.23system 1:26.68elapsed 18%CPU (0avgtext+0avgdata 255688maxresident)k ``` Reviewed By: hx235 Differential Revision: D30048013 Pulled By: ajkr fbshipit-source-id: 6741bba9d9dfbccab359806d725105817fef818b
3 years ago
// Whichever thread reaches here first performs duty (1) as described
// above.
int64_t wait_until = options_.clock->NowMicros() + time_until_refill_us;
RecordTick(stats, NUMBER_RATE_LIMITER_DRAINS);
++num_drains_;
Simplify GenericRateLimiter algorithm (#8602) Summary: `GenericRateLimiter` slow path handles requests that cannot be satisfied immediately. Such requests enter a queue, and their thread stays in `Request()` until they are granted or the rate limiter is stopped. These threads are responsible for unblocking themselves. The work to do so is split into two main duties. (1) Waiting for the next refill time. (2) Refilling the bytes and granting requests. Prior to this PR, the slow path logic involved a leader election algorithm to pick one thread to perform (1) followed by (2). It elected the thread whose request was at the front of the highest priority non-empty queue since that request was most likely to be granted. This algorithm was efficient in terms of reducing intermediate wakeups, which is a thread waking up only to resume waiting after finding its request is not granted. However, the conceptual complexity of this algorithm was too high. It took me a long time to draw a timeline to understand how it works for just one edge case yet there were so many. This PR drops the leader election to reduce conceptual complexity. Now, the two duties can be performed by whichever thread acquires the lock first. The risk of this change is increasing the number of intermediate wakeups, however, we took steps to mitigate that. - `wait_until_refill_pending_` flag ensures only one thread performs (1). This\ prevents the thundering herd problem at the next refill time. The remaining\ threads wait on their condition variable with an unbounded duration -- thus we\ must remember to notify them to ensure forward progress. - (1) is typically done by a thread at the front of a queue. This is trivial\ when the queues are initially empty as the first choice that arrives must be\ the only entry in its queue. When queues are initially non-empty, we achieve\ this by having (2) notify a thread at the front of a queue (preferring higher\ priority) to perform the next duty. - We do not require any additional wakeup for (2). Typically it will just be\ done by the thread that finished (1). Combined, the second and third bullet points above suggest the refill/granting will typically be done by a request at the front of its queue. This is important because one wakeup is saved when a granted request happens to be in an already running thread. Note there are a few cases that still lead to intermediate wakeup, however. The first two are existing issues that also apply to the old algorithm, however, the third (including both subpoints) is new. - No request may be granted (only possible when rate limit dynamically\ decreases). - Requests from a different queue may be granted. - (2) may be run by a non-front request thread causing it to not be granted even\ if some requests in that same queue are granted. It can happen for a couple\ (unlikely) reasons. - A new request may sneak in and grab the lock at the refill time, before the\ thread finishing (1) can wake up and grab it. - A new request may sneak in and grab the lock and execute (1) before (2)'s\ chosen candidate can wake up and grab the lock. Then that non-front request\ thread performing (1) can carry over to perform (2). Pull Request resolved: https://github.com/facebook/rocksdb/pull/8602 Test Plan: - Use existing tests. The edge cases listed in the comment are all performance\ related; I could not really think of any related to correctness. The logic\ looks the same whether a thread wakes up/finishes its work early/on-time/late,\ or whether the thread is chosen vs. "steals" the work. - Verified write throughput and CPU overhead are basically the same with and\ without this change, even in a rate limiter heavy workload: Test command: ``` $ rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -benchmarks=fillrandom -num_multi_db=64 -num_low_pri_threads=64 -num_high_pri_threads=64 -write_buffer_size=262144 -target_file_size_base=262144 -max_bytes_for_level_base=1048576 -rate_limiter_bytes_per_sec=16777216 -key_size=24 -value_size=1000 -num=10000 -compression_type=none -rate_limiter_refill_period_us=1000 ``` Results before this PR: ``` fillrandom : 108.463 micros/op 9219 ops/sec; 9.0 MB/s 7.40user 8.84system 1:26.20elapsed 18%CPU (0avgtext+0avgdata 256140maxresident)k ``` Results after this PR: ``` fillrandom : 108.108 micros/op 9250 ops/sec; 9.0 MB/s 7.45user 8.23system 1:26.68elapsed 18%CPU (0avgtext+0avgdata 255688maxresident)k ``` Reviewed By: hx235 Differential Revision: D30048013 Pulled By: ajkr fbshipit-source-id: 6741bba9d9dfbccab359806d725105817fef818b
3 years ago
wait_until_refill_pending_ = true;
r.cv.TimedWait(wait_until);
TEST_SYNC_POINT_CALLBACK("GenericRateLimiter::Request:PostTimedWait",
&time_until_refill_us);
wait_until_refill_pending_ = false;
}
} else {
Simplify GenericRateLimiter algorithm (#8602) Summary: `GenericRateLimiter` slow path handles requests that cannot be satisfied immediately. Such requests enter a queue, and their thread stays in `Request()` until they are granted or the rate limiter is stopped. These threads are responsible for unblocking themselves. The work to do so is split into two main duties. (1) Waiting for the next refill time. (2) Refilling the bytes and granting requests. Prior to this PR, the slow path logic involved a leader election algorithm to pick one thread to perform (1) followed by (2). It elected the thread whose request was at the front of the highest priority non-empty queue since that request was most likely to be granted. This algorithm was efficient in terms of reducing intermediate wakeups, which is a thread waking up only to resume waiting after finding its request is not granted. However, the conceptual complexity of this algorithm was too high. It took me a long time to draw a timeline to understand how it works for just one edge case yet there were so many. This PR drops the leader election to reduce conceptual complexity. Now, the two duties can be performed by whichever thread acquires the lock first. The risk of this change is increasing the number of intermediate wakeups, however, we took steps to mitigate that. - `wait_until_refill_pending_` flag ensures only one thread performs (1). This\ prevents the thundering herd problem at the next refill time. The remaining\ threads wait on their condition variable with an unbounded duration -- thus we\ must remember to notify them to ensure forward progress. - (1) is typically done by a thread at the front of a queue. This is trivial\ when the queues are initially empty as the first choice that arrives must be\ the only entry in its queue. When queues are initially non-empty, we achieve\ this by having (2) notify a thread at the front of a queue (preferring higher\ priority) to perform the next duty. - We do not require any additional wakeup for (2). Typically it will just be\ done by the thread that finished (1). Combined, the second and third bullet points above suggest the refill/granting will typically be done by a request at the front of its queue. This is important because one wakeup is saved when a granted request happens to be in an already running thread. Note there are a few cases that still lead to intermediate wakeup, however. The first two are existing issues that also apply to the old algorithm, however, the third (including both subpoints) is new. - No request may be granted (only possible when rate limit dynamically\ decreases). - Requests from a different queue may be granted. - (2) may be run by a non-front request thread causing it to not be granted even\ if some requests in that same queue are granted. It can happen for a couple\ (unlikely) reasons. - A new request may sneak in and grab the lock at the refill time, before the\ thread finishing (1) can wake up and grab it. - A new request may sneak in and grab the lock and execute (1) before (2)'s\ chosen candidate can wake up and grab the lock. Then that non-front request\ thread performing (1) can carry over to perform (2). Pull Request resolved: https://github.com/facebook/rocksdb/pull/8602 Test Plan: - Use existing tests. The edge cases listed in the comment are all performance\ related; I could not really think of any related to correctness. The logic\ looks the same whether a thread wakes up/finishes its work early/on-time/late,\ or whether the thread is chosen vs. "steals" the work. - Verified write throughput and CPU overhead are basically the same with and\ without this change, even in a rate limiter heavy workload: Test command: ``` $ rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -benchmarks=fillrandom -num_multi_db=64 -num_low_pri_threads=64 -num_high_pri_threads=64 -write_buffer_size=262144 -target_file_size_base=262144 -max_bytes_for_level_base=1048576 -rate_limiter_bytes_per_sec=16777216 -key_size=24 -value_size=1000 -num=10000 -compression_type=none -rate_limiter_refill_period_us=1000 ``` Results before this PR: ``` fillrandom : 108.463 micros/op 9219 ops/sec; 9.0 MB/s 7.40user 8.84system 1:26.20elapsed 18%CPU (0avgtext+0avgdata 256140maxresident)k ``` Results after this PR: ``` fillrandom : 108.108 micros/op 9250 ops/sec; 9.0 MB/s 7.45user 8.23system 1:26.68elapsed 18%CPU (0avgtext+0avgdata 255688maxresident)k ``` Reviewed By: hx235 Differential Revision: D30048013 Pulled By: ajkr fbshipit-source-id: 6741bba9d9dfbccab359806d725105817fef818b
3 years ago
// Whichever thread reaches here first performs duty (2) as described
// above.
RefillBytesAndGrantRequests();
if (r.granted) {
// If there is any remaining requests, make sure there exists at least
// one candidate is awake for future duties by signaling a front request
// of a queue.
Implement superior user & mid IO priority level in GenericRateLimiter (#8595) Summary: Context: An extra IO_USER priority in rate limiter allows users to optionally charge WAL writes / SST reads to rate limiter at this priority level, which then has higher priority than IO_HIGH and IO_LOW. With an extra IO_USER priority, it allows users to better specify the relative urgency/importance among different requests in rate limiter. As a consequence, IO resource management can better prioritize and limit resource based on user's need. The IO_USER is implemented as superior priority in GenericRateLimiter, in the sense that its request queue will always be iterated first without being constrained to fairness. The reason is that the notion of fairness is only meaningful in helping lower priorities in background IO (i.e, IO_HIGH/MID/LOW) to gain some fair chance to run so that it does not block foreground IO (i.e, the ones that are charged at the level of IO_USER). As we can see, the ultimate goal here is to not blocking foreground IO at IO_USER level, which justifies the superiority of IO_USER. Similar benefits exist for IO_MID priority. - Rewrote the logic of deciding the order of iterating request queues of high/low priorities to include the extra user/mid priority w/o affecting the existing behavior (see PR's [comment](https://github.com/facebook/rocksdb/pull/8595/files#r678749331)) - Included the request queue of user-pri/mid-pri in the code path of next-leader-candidate signaling and GenericRateLimiter's destructor - Included the extra user/mid-pri in bookkeeping data structures: total_bytes_through_ and total_requests_ - Re-written the previous impl of explicitly iterating priorities with a loop from Env::IO_LOW to Env::IO_TOTAL Pull Request resolved: https://github.com/facebook/rocksdb/pull/8595 Test Plan: - passed existing rate_limiter_test.cc - passed added unit tests in rate_limiter_test.cc - run performance test to verify performance with only high/low requests is not affected by this change - Set-up command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom --duration=5 --compression_type=none --num=100000000 --disable_auto_compactions=true --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1))` - Test command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=overwrite --use_existing_db=true --disable_wal=true --duration=30 --compression_type=none --num=100000000 --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1)) --statistics=true --rate_limiter_bytes_per_sec=1048576 --rate_limiter_refill_period_us=1000 --threads=32 |& grep -E '(flush|compact)\.write\.bytes'` - Before (on branch upstream/master): `rocksdb.compact.write.bytes COUNT : 4014162` `rocksdb.flush.write.bytes COUNT : 26715832` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.66 - After (on branch rate_limiter_user_pri): `rocksdb.compact.write.bytes COUNT : 3807822` `rocksdb.flush.write.bytes COUNT : 26098659` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.85 Reviewed By: ajkr Differential Revision: D30577783 Pulled By: hx235 fbshipit-source-id: 0881f2705ffd13ecd331256bde7e8ec874a353f4
3 years ago
for (int i = Env::IO_TOTAL - 1; i >= Env::IO_LOW; --i) {
std::deque<Req*> queue = queue_[i];
if (!queue.empty()) {
queue.front()->cv.Signal();
break;
}
}
}
Simplify GenericRateLimiter algorithm (#8602) Summary: `GenericRateLimiter` slow path handles requests that cannot be satisfied immediately. Such requests enter a queue, and their thread stays in `Request()` until they are granted or the rate limiter is stopped. These threads are responsible for unblocking themselves. The work to do so is split into two main duties. (1) Waiting for the next refill time. (2) Refilling the bytes and granting requests. Prior to this PR, the slow path logic involved a leader election algorithm to pick one thread to perform (1) followed by (2). It elected the thread whose request was at the front of the highest priority non-empty queue since that request was most likely to be granted. This algorithm was efficient in terms of reducing intermediate wakeups, which is a thread waking up only to resume waiting after finding its request is not granted. However, the conceptual complexity of this algorithm was too high. It took me a long time to draw a timeline to understand how it works for just one edge case yet there were so many. This PR drops the leader election to reduce conceptual complexity. Now, the two duties can be performed by whichever thread acquires the lock first. The risk of this change is increasing the number of intermediate wakeups, however, we took steps to mitigate that. - `wait_until_refill_pending_` flag ensures only one thread performs (1). This\ prevents the thundering herd problem at the next refill time. The remaining\ threads wait on their condition variable with an unbounded duration -- thus we\ must remember to notify them to ensure forward progress. - (1) is typically done by a thread at the front of a queue. This is trivial\ when the queues are initially empty as the first choice that arrives must be\ the only entry in its queue. When queues are initially non-empty, we achieve\ this by having (2) notify a thread at the front of a queue (preferring higher\ priority) to perform the next duty. - We do not require any additional wakeup for (2). Typically it will just be\ done by the thread that finished (1). Combined, the second and third bullet points above suggest the refill/granting will typically be done by a request at the front of its queue. This is important because one wakeup is saved when a granted request happens to be in an already running thread. Note there are a few cases that still lead to intermediate wakeup, however. The first two are existing issues that also apply to the old algorithm, however, the third (including both subpoints) is new. - No request may be granted (only possible when rate limit dynamically\ decreases). - Requests from a different queue may be granted. - (2) may be run by a non-front request thread causing it to not be granted even\ if some requests in that same queue are granted. It can happen for a couple\ (unlikely) reasons. - A new request may sneak in and grab the lock at the refill time, before the\ thread finishing (1) can wake up and grab it. - A new request may sneak in and grab the lock and execute (1) before (2)'s\ chosen candidate can wake up and grab the lock. Then that non-front request\ thread performing (1) can carry over to perform (2). Pull Request resolved: https://github.com/facebook/rocksdb/pull/8602 Test Plan: - Use existing tests. The edge cases listed in the comment are all performance\ related; I could not really think of any related to correctness. The logic\ looks the same whether a thread wakes up/finishes its work early/on-time/late,\ or whether the thread is chosen vs. "steals" the work. - Verified write throughput and CPU overhead are basically the same with and\ without this change, even in a rate limiter heavy workload: Test command: ``` $ rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -benchmarks=fillrandom -num_multi_db=64 -num_low_pri_threads=64 -num_high_pri_threads=64 -write_buffer_size=262144 -target_file_size_base=262144 -max_bytes_for_level_base=1048576 -rate_limiter_bytes_per_sec=16777216 -key_size=24 -value_size=1000 -num=10000 -compression_type=none -rate_limiter_refill_period_us=1000 ``` Results before this PR: ``` fillrandom : 108.463 micros/op 9219 ops/sec; 9.0 MB/s 7.40user 8.84system 1:26.20elapsed 18%CPU (0avgtext+0avgdata 256140maxresident)k ``` Results after this PR: ``` fillrandom : 108.108 micros/op 9250 ops/sec; 9.0 MB/s 7.45user 8.23system 1:26.68elapsed 18%CPU (0avgtext+0avgdata 255688maxresident)k ``` Reviewed By: hx235 Differential Revision: D30048013 Pulled By: ajkr fbshipit-source-id: 6741bba9d9dfbccab359806d725105817fef818b
3 years ago
}
// Invariant: non-granted request is always in one queue, and granted
// request is always in zero queues.
#ifndef NDEBUG
int num_found = 0;
for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
if (std::find(queue_[i].begin(), queue_[i].end(), &r) !=
queue_[i].end()) {
++num_found;
}
}
if (r.granted) {
assert(num_found == 0);
} else {
Simplify GenericRateLimiter algorithm (#8602) Summary: `GenericRateLimiter` slow path handles requests that cannot be satisfied immediately. Such requests enter a queue, and their thread stays in `Request()` until they are granted or the rate limiter is stopped. These threads are responsible for unblocking themselves. The work to do so is split into two main duties. (1) Waiting for the next refill time. (2) Refilling the bytes and granting requests. Prior to this PR, the slow path logic involved a leader election algorithm to pick one thread to perform (1) followed by (2). It elected the thread whose request was at the front of the highest priority non-empty queue since that request was most likely to be granted. This algorithm was efficient in terms of reducing intermediate wakeups, which is a thread waking up only to resume waiting after finding its request is not granted. However, the conceptual complexity of this algorithm was too high. It took me a long time to draw a timeline to understand how it works for just one edge case yet there were so many. This PR drops the leader election to reduce conceptual complexity. Now, the two duties can be performed by whichever thread acquires the lock first. The risk of this change is increasing the number of intermediate wakeups, however, we took steps to mitigate that. - `wait_until_refill_pending_` flag ensures only one thread performs (1). This\ prevents the thundering herd problem at the next refill time. The remaining\ threads wait on their condition variable with an unbounded duration -- thus we\ must remember to notify them to ensure forward progress. - (1) is typically done by a thread at the front of a queue. This is trivial\ when the queues are initially empty as the first choice that arrives must be\ the only entry in its queue. When queues are initially non-empty, we achieve\ this by having (2) notify a thread at the front of a queue (preferring higher\ priority) to perform the next duty. - We do not require any additional wakeup for (2). Typically it will just be\ done by the thread that finished (1). Combined, the second and third bullet points above suggest the refill/granting will typically be done by a request at the front of its queue. This is important because one wakeup is saved when a granted request happens to be in an already running thread. Note there are a few cases that still lead to intermediate wakeup, however. The first two are existing issues that also apply to the old algorithm, however, the third (including both subpoints) is new. - No request may be granted (only possible when rate limit dynamically\ decreases). - Requests from a different queue may be granted. - (2) may be run by a non-front request thread causing it to not be granted even\ if some requests in that same queue are granted. It can happen for a couple\ (unlikely) reasons. - A new request may sneak in and grab the lock at the refill time, before the\ thread finishing (1) can wake up and grab it. - A new request may sneak in and grab the lock and execute (1) before (2)'s\ chosen candidate can wake up and grab the lock. Then that non-front request\ thread performing (1) can carry over to perform (2). Pull Request resolved: https://github.com/facebook/rocksdb/pull/8602 Test Plan: - Use existing tests. The edge cases listed in the comment are all performance\ related; I could not really think of any related to correctness. The logic\ looks the same whether a thread wakes up/finishes its work early/on-time/late,\ or whether the thread is chosen vs. "steals" the work. - Verified write throughput and CPU overhead are basically the same with and\ without this change, even in a rate limiter heavy workload: Test command: ``` $ rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -benchmarks=fillrandom -num_multi_db=64 -num_low_pri_threads=64 -num_high_pri_threads=64 -write_buffer_size=262144 -target_file_size_base=262144 -max_bytes_for_level_base=1048576 -rate_limiter_bytes_per_sec=16777216 -key_size=24 -value_size=1000 -num=10000 -compression_type=none -rate_limiter_refill_period_us=1000 ``` Results before this PR: ``` fillrandom : 108.463 micros/op 9219 ops/sec; 9.0 MB/s 7.40user 8.84system 1:26.20elapsed 18%CPU (0avgtext+0avgdata 256140maxresident)k ``` Results after this PR: ``` fillrandom : 108.108 micros/op 9250 ops/sec; 9.0 MB/s 7.45user 8.23system 1:26.68elapsed 18%CPU (0avgtext+0avgdata 255688maxresident)k ``` Reviewed By: hx235 Differential Revision: D30048013 Pulled By: ajkr fbshipit-source-id: 6741bba9d9dfbccab359806d725105817fef818b
3 years ago
assert(num_found == 1);
}
Simplify GenericRateLimiter algorithm (#8602) Summary: `GenericRateLimiter` slow path handles requests that cannot be satisfied immediately. Such requests enter a queue, and their thread stays in `Request()` until they are granted or the rate limiter is stopped. These threads are responsible for unblocking themselves. The work to do so is split into two main duties. (1) Waiting for the next refill time. (2) Refilling the bytes and granting requests. Prior to this PR, the slow path logic involved a leader election algorithm to pick one thread to perform (1) followed by (2). It elected the thread whose request was at the front of the highest priority non-empty queue since that request was most likely to be granted. This algorithm was efficient in terms of reducing intermediate wakeups, which is a thread waking up only to resume waiting after finding its request is not granted. However, the conceptual complexity of this algorithm was too high. It took me a long time to draw a timeline to understand how it works for just one edge case yet there were so many. This PR drops the leader election to reduce conceptual complexity. Now, the two duties can be performed by whichever thread acquires the lock first. The risk of this change is increasing the number of intermediate wakeups, however, we took steps to mitigate that. - `wait_until_refill_pending_` flag ensures only one thread performs (1). This\ prevents the thundering herd problem at the next refill time. The remaining\ threads wait on their condition variable with an unbounded duration -- thus we\ must remember to notify them to ensure forward progress. - (1) is typically done by a thread at the front of a queue. This is trivial\ when the queues are initially empty as the first choice that arrives must be\ the only entry in its queue. When queues are initially non-empty, we achieve\ this by having (2) notify a thread at the front of a queue (preferring higher\ priority) to perform the next duty. - We do not require any additional wakeup for (2). Typically it will just be\ done by the thread that finished (1). Combined, the second and third bullet points above suggest the refill/granting will typically be done by a request at the front of its queue. This is important because one wakeup is saved when a granted request happens to be in an already running thread. Note there are a few cases that still lead to intermediate wakeup, however. The first two are existing issues that also apply to the old algorithm, however, the third (including both subpoints) is new. - No request may be granted (only possible when rate limit dynamically\ decreases). - Requests from a different queue may be granted. - (2) may be run by a non-front request thread causing it to not be granted even\ if some requests in that same queue are granted. It can happen for a couple\ (unlikely) reasons. - A new request may sneak in and grab the lock at the refill time, before the\ thread finishing (1) can wake up and grab it. - A new request may sneak in and grab the lock and execute (1) before (2)'s\ chosen candidate can wake up and grab the lock. Then that non-front request\ thread performing (1) can carry over to perform (2). Pull Request resolved: https://github.com/facebook/rocksdb/pull/8602 Test Plan: - Use existing tests. The edge cases listed in the comment are all performance\ related; I could not really think of any related to correctness. The logic\ looks the same whether a thread wakes up/finishes its work early/on-time/late,\ or whether the thread is chosen vs. "steals" the work. - Verified write throughput and CPU overhead are basically the same with and\ without this change, even in a rate limiter heavy workload: Test command: ``` $ rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -benchmarks=fillrandom -num_multi_db=64 -num_low_pri_threads=64 -num_high_pri_threads=64 -write_buffer_size=262144 -target_file_size_base=262144 -max_bytes_for_level_base=1048576 -rate_limiter_bytes_per_sec=16777216 -key_size=24 -value_size=1000 -num=10000 -compression_type=none -rate_limiter_refill_period_us=1000 ``` Results before this PR: ``` fillrandom : 108.463 micros/op 9219 ops/sec; 9.0 MB/s 7.40user 8.84system 1:26.20elapsed 18%CPU (0avgtext+0avgdata 256140maxresident)k ``` Results after this PR: ``` fillrandom : 108.108 micros/op 9250 ops/sec; 9.0 MB/s 7.45user 8.23system 1:26.68elapsed 18%CPU (0avgtext+0avgdata 255688maxresident)k ``` Reviewed By: hx235 Differential Revision: D30048013 Pulled By: ajkr fbshipit-source-id: 6741bba9d9dfbccab359806d725105817fef818b
3 years ago
#endif // NDEBUG
} while (!stop_ && !r.granted);
if (stop_) {
// It is now in the clean-up of ~GenericRateLimiter().
// Therefore any woken-up request will have come out of the loop and then
// exit here. It might or might not have been satisfied.
--requests_to_wait_;
exit_cv_.Signal();
}
}
Implement superior user & mid IO priority level in GenericRateLimiter (#8595) Summary: Context: An extra IO_USER priority in rate limiter allows users to optionally charge WAL writes / SST reads to rate limiter at this priority level, which then has higher priority than IO_HIGH and IO_LOW. With an extra IO_USER priority, it allows users to better specify the relative urgency/importance among different requests in rate limiter. As a consequence, IO resource management can better prioritize and limit resource based on user's need. The IO_USER is implemented as superior priority in GenericRateLimiter, in the sense that its request queue will always be iterated first without being constrained to fairness. The reason is that the notion of fairness is only meaningful in helping lower priorities in background IO (i.e, IO_HIGH/MID/LOW) to gain some fair chance to run so that it does not block foreground IO (i.e, the ones that are charged at the level of IO_USER). As we can see, the ultimate goal here is to not blocking foreground IO at IO_USER level, which justifies the superiority of IO_USER. Similar benefits exist for IO_MID priority. - Rewrote the logic of deciding the order of iterating request queues of high/low priorities to include the extra user/mid priority w/o affecting the existing behavior (see PR's [comment](https://github.com/facebook/rocksdb/pull/8595/files#r678749331)) - Included the request queue of user-pri/mid-pri in the code path of next-leader-candidate signaling and GenericRateLimiter's destructor - Included the extra user/mid-pri in bookkeeping data structures: total_bytes_through_ and total_requests_ - Re-written the previous impl of explicitly iterating priorities with a loop from Env::IO_LOW to Env::IO_TOTAL Pull Request resolved: https://github.com/facebook/rocksdb/pull/8595 Test Plan: - passed existing rate_limiter_test.cc - passed added unit tests in rate_limiter_test.cc - run performance test to verify performance with only high/low requests is not affected by this change - Set-up command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom --duration=5 --compression_type=none --num=100000000 --disable_auto_compactions=true --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1))` - Test command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=overwrite --use_existing_db=true --disable_wal=true --duration=30 --compression_type=none --num=100000000 --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1)) --statistics=true --rate_limiter_bytes_per_sec=1048576 --rate_limiter_refill_period_us=1000 --threads=32 |& grep -E '(flush|compact)\.write\.bytes'` - Before (on branch upstream/master): `rocksdb.compact.write.bytes COUNT : 4014162` `rocksdb.flush.write.bytes COUNT : 26715832` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.66 - After (on branch rate_limiter_user_pri): `rocksdb.compact.write.bytes COUNT : 3807822` `rocksdb.flush.write.bytes COUNT : 26098659` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.85 Reviewed By: ajkr Differential Revision: D30577783 Pulled By: hx235 fbshipit-source-id: 0881f2705ffd13ecd331256bde7e8ec874a353f4
3 years ago
std::vector<Env::IOPriority>
GenericRateLimiter::GeneratePriorityIterationOrder() {
std::vector<Env::IOPriority> pri_iteration_order(Env::IO_TOTAL /* 4 */);
// We make Env::IO_USER a superior priority by always iterating its queue
// first
pri_iteration_order[0] = Env::IO_USER;
bool high_pri_iterated_after_mid_low_pri = rnd_.OneIn(options_.fairness);
Implement superior user & mid IO priority level in GenericRateLimiter (#8595) Summary: Context: An extra IO_USER priority in rate limiter allows users to optionally charge WAL writes / SST reads to rate limiter at this priority level, which then has higher priority than IO_HIGH and IO_LOW. With an extra IO_USER priority, it allows users to better specify the relative urgency/importance among different requests in rate limiter. As a consequence, IO resource management can better prioritize and limit resource based on user's need. The IO_USER is implemented as superior priority in GenericRateLimiter, in the sense that its request queue will always be iterated first without being constrained to fairness. The reason is that the notion of fairness is only meaningful in helping lower priorities in background IO (i.e, IO_HIGH/MID/LOW) to gain some fair chance to run so that it does not block foreground IO (i.e, the ones that are charged at the level of IO_USER). As we can see, the ultimate goal here is to not blocking foreground IO at IO_USER level, which justifies the superiority of IO_USER. Similar benefits exist for IO_MID priority. - Rewrote the logic of deciding the order of iterating request queues of high/low priorities to include the extra user/mid priority w/o affecting the existing behavior (see PR's [comment](https://github.com/facebook/rocksdb/pull/8595/files#r678749331)) - Included the request queue of user-pri/mid-pri in the code path of next-leader-candidate signaling and GenericRateLimiter's destructor - Included the extra user/mid-pri in bookkeeping data structures: total_bytes_through_ and total_requests_ - Re-written the previous impl of explicitly iterating priorities with a loop from Env::IO_LOW to Env::IO_TOTAL Pull Request resolved: https://github.com/facebook/rocksdb/pull/8595 Test Plan: - passed existing rate_limiter_test.cc - passed added unit tests in rate_limiter_test.cc - run performance test to verify performance with only high/low requests is not affected by this change - Set-up command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom --duration=5 --compression_type=none --num=100000000 --disable_auto_compactions=true --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1))` - Test command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=overwrite --use_existing_db=true --disable_wal=true --duration=30 --compression_type=none --num=100000000 --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1)) --statistics=true --rate_limiter_bytes_per_sec=1048576 --rate_limiter_refill_period_us=1000 --threads=32 |& grep -E '(flush|compact)\.write\.bytes'` - Before (on branch upstream/master): `rocksdb.compact.write.bytes COUNT : 4014162` `rocksdb.flush.write.bytes COUNT : 26715832` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.66 - After (on branch rate_limiter_user_pri): `rocksdb.compact.write.bytes COUNT : 3807822` `rocksdb.flush.write.bytes COUNT : 26098659` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.85 Reviewed By: ajkr Differential Revision: D30577783 Pulled By: hx235 fbshipit-source-id: 0881f2705ffd13ecd331256bde7e8ec874a353f4
3 years ago
TEST_SYNC_POINT_CALLBACK(
"GenericRateLimiter::GeneratePriorityIterationOrder::"
"PostRandomOneInFairnessForHighPri",
&high_pri_iterated_after_mid_low_pri);
bool mid_pri_itereated_after_low_pri = rnd_.OneIn(options_.fairness);
Implement superior user & mid IO priority level in GenericRateLimiter (#8595) Summary: Context: An extra IO_USER priority in rate limiter allows users to optionally charge WAL writes / SST reads to rate limiter at this priority level, which then has higher priority than IO_HIGH and IO_LOW. With an extra IO_USER priority, it allows users to better specify the relative urgency/importance among different requests in rate limiter. As a consequence, IO resource management can better prioritize and limit resource based on user's need. The IO_USER is implemented as superior priority in GenericRateLimiter, in the sense that its request queue will always be iterated first without being constrained to fairness. The reason is that the notion of fairness is only meaningful in helping lower priorities in background IO (i.e, IO_HIGH/MID/LOW) to gain some fair chance to run so that it does not block foreground IO (i.e, the ones that are charged at the level of IO_USER). As we can see, the ultimate goal here is to not blocking foreground IO at IO_USER level, which justifies the superiority of IO_USER. Similar benefits exist for IO_MID priority. - Rewrote the logic of deciding the order of iterating request queues of high/low priorities to include the extra user/mid priority w/o affecting the existing behavior (see PR's [comment](https://github.com/facebook/rocksdb/pull/8595/files#r678749331)) - Included the request queue of user-pri/mid-pri in the code path of next-leader-candidate signaling and GenericRateLimiter's destructor - Included the extra user/mid-pri in bookkeeping data structures: total_bytes_through_ and total_requests_ - Re-written the previous impl of explicitly iterating priorities with a loop from Env::IO_LOW to Env::IO_TOTAL Pull Request resolved: https://github.com/facebook/rocksdb/pull/8595 Test Plan: - passed existing rate_limiter_test.cc - passed added unit tests in rate_limiter_test.cc - run performance test to verify performance with only high/low requests is not affected by this change - Set-up command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom --duration=5 --compression_type=none --num=100000000 --disable_auto_compactions=true --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1))` - Test command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=overwrite --use_existing_db=true --disable_wal=true --duration=30 --compression_type=none --num=100000000 --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1)) --statistics=true --rate_limiter_bytes_per_sec=1048576 --rate_limiter_refill_period_us=1000 --threads=32 |& grep -E '(flush|compact)\.write\.bytes'` - Before (on branch upstream/master): `rocksdb.compact.write.bytes COUNT : 4014162` `rocksdb.flush.write.bytes COUNT : 26715832` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.66 - After (on branch rate_limiter_user_pri): `rocksdb.compact.write.bytes COUNT : 3807822` `rocksdb.flush.write.bytes COUNT : 26098659` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.85 Reviewed By: ajkr Differential Revision: D30577783 Pulled By: hx235 fbshipit-source-id: 0881f2705ffd13ecd331256bde7e8ec874a353f4
3 years ago
TEST_SYNC_POINT_CALLBACK(
"GenericRateLimiter::GeneratePriorityIterationOrder::"
"PostRandomOneInFairnessForMidPri",
&mid_pri_itereated_after_low_pri);
if (high_pri_iterated_after_mid_low_pri) {
pri_iteration_order[3] = Env::IO_HIGH;
pri_iteration_order[2] =
mid_pri_itereated_after_low_pri ? Env::IO_MID : Env::IO_LOW;
pri_iteration_order[1] =
(pri_iteration_order[2] == Env::IO_MID) ? Env::IO_LOW : Env::IO_MID;
} else {
pri_iteration_order[1] = Env::IO_HIGH;
pri_iteration_order[3] =
mid_pri_itereated_after_low_pri ? Env::IO_MID : Env::IO_LOW;
pri_iteration_order[2] =
(pri_iteration_order[3] == Env::IO_MID) ? Env::IO_LOW : Env::IO_MID;
}
TEST_SYNC_POINT_CALLBACK(
"GenericRateLimiter::GeneratePriorityIterationOrder::"
"PreReturnPriIterationOrder",
&pri_iteration_order);
return pri_iteration_order;
}
void GenericRateLimiter::RefillBytesAndGrantRequests() {
TEST_SYNC_POINT("GenericRateLimiter::RefillBytesAndGrantRequests");
next_refill_us_ = NowMicrosMonotonic() + options_.refill_period_us;
// Carry over the left over quota from the last period
auto refill_bytes_per_period =
refill_bytes_per_period_.load(std::memory_order_relaxed);
if (available_bytes_ < refill_bytes_per_period) {
available_bytes_ += refill_bytes_per_period;
}
Implement superior user & mid IO priority level in GenericRateLimiter (#8595) Summary: Context: An extra IO_USER priority in rate limiter allows users to optionally charge WAL writes / SST reads to rate limiter at this priority level, which then has higher priority than IO_HIGH and IO_LOW. With an extra IO_USER priority, it allows users to better specify the relative urgency/importance among different requests in rate limiter. As a consequence, IO resource management can better prioritize and limit resource based on user's need. The IO_USER is implemented as superior priority in GenericRateLimiter, in the sense that its request queue will always be iterated first without being constrained to fairness. The reason is that the notion of fairness is only meaningful in helping lower priorities in background IO (i.e, IO_HIGH/MID/LOW) to gain some fair chance to run so that it does not block foreground IO (i.e, the ones that are charged at the level of IO_USER). As we can see, the ultimate goal here is to not blocking foreground IO at IO_USER level, which justifies the superiority of IO_USER. Similar benefits exist for IO_MID priority. - Rewrote the logic of deciding the order of iterating request queues of high/low priorities to include the extra user/mid priority w/o affecting the existing behavior (see PR's [comment](https://github.com/facebook/rocksdb/pull/8595/files#r678749331)) - Included the request queue of user-pri/mid-pri in the code path of next-leader-candidate signaling and GenericRateLimiter's destructor - Included the extra user/mid-pri in bookkeeping data structures: total_bytes_through_ and total_requests_ - Re-written the previous impl of explicitly iterating priorities with a loop from Env::IO_LOW to Env::IO_TOTAL Pull Request resolved: https://github.com/facebook/rocksdb/pull/8595 Test Plan: - passed existing rate_limiter_test.cc - passed added unit tests in rate_limiter_test.cc - run performance test to verify performance with only high/low requests is not affected by this change - Set-up command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom --duration=5 --compression_type=none --num=100000000 --disable_auto_compactions=true --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1))` - Test command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=overwrite --use_existing_db=true --disable_wal=true --duration=30 --compression_type=none --num=100000000 --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1)) --statistics=true --rate_limiter_bytes_per_sec=1048576 --rate_limiter_refill_period_us=1000 --threads=32 |& grep -E '(flush|compact)\.write\.bytes'` - Before (on branch upstream/master): `rocksdb.compact.write.bytes COUNT : 4014162` `rocksdb.flush.write.bytes COUNT : 26715832` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.66 - After (on branch rate_limiter_user_pri): `rocksdb.compact.write.bytes COUNT : 3807822` `rocksdb.flush.write.bytes COUNT : 26098659` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.85 Reviewed By: ajkr Differential Revision: D30577783 Pulled By: hx235 fbshipit-source-id: 0881f2705ffd13ecd331256bde7e8ec874a353f4
3 years ago
std::vector<Env::IOPriority> pri_iteration_order =
GeneratePriorityIterationOrder();
for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
assert(!pri_iteration_order.empty());
Env::IOPriority current_pri = pri_iteration_order[i];
auto* queue = &queue_[current_pri];
while (!queue->empty()) {
auto* next_req = queue->front();
fix rate limiter to avoid starvation Summary: The current implementation of rate limiter has the possibility to introduce resource starvation when change its limit. This diff aims to fix this problem by consuming request bytes partially. Test Plan: ``` ./rate_limiter_test [==========] Running 4 tests from 1 test case. [----------] Global test environment set-up. [----------] 4 tests from RateLimiterTest [ RUN ] RateLimiterTest.OverflowRate [ OK ] RateLimiterTest.OverflowRate (0 ms) [ RUN ] RateLimiterTest.StartStop [ OK ] RateLimiterTest.StartStop (0 ms) [ RUN ] RateLimiterTest.Rate request size [1 - 1023], limit 10 KB/sec, actual rate: 10.355712 KB/sec, elapsed 2.00 seconds request size [1 - 1023], limit 20 KB/sec, actual rate: 19.136564 KB/sec, elapsed 2.00 seconds request size [1 - 2047], limit 20 KB/sec, actual rate: 20.783976 KB/sec, elapsed 2.10 seconds request size [1 - 2047], limit 40 KB/sec, actual rate: 39.308144 KB/sec, elapsed 2.10 seconds request size [1 - 4095], limit 40 KB/sec, actual rate: 40.318349 KB/sec, elapsed 2.20 seconds request size [1 - 4095], limit 80 KB/sec, actual rate: 79.667396 KB/sec, elapsed 2.20 seconds request size [1 - 8191], limit 80 KB/sec, actual rate: 81.807158 KB/sec, elapsed 2.30 seconds request size [1 - 8191], limit 160 KB/sec, actual rate: 160.659761 KB/sec, elapsed 2.20 seconds request size [1 - 16383], limit 160 KB/sec, actual rate: 160.700990 KB/sec, elapsed 3.00 seconds request size [1 - 16383], limit 320 KB/sec, actual rate: 317.639481 KB/sec, elapsed 2.50 seconds [ OK ] RateLimiterTest.Rate (22618 ms) [ RUN ] RateLimiterTest.LimitChangeTest [COMPLETE] request size 10 KB, new limit 20KB/sec, refill period 1000 ms [COMPLETE] request size 10 KB, new limit 5KB/sec, refill period 1000 ms [COMPLETE] request size 20 KB, new limit 40KB/sec, refill period 1000 ms [COMPLETE] request size 20 KB, new limit 10KB/sec, refill period 1000 ms [COMPLETE] request size 40 KB, new limit 80KB/sec, refill period 1000 ms [COMPLETE] request size 40 KB, new limit 20KB/sec, refill period 1000 ms [COMPLETE] request size 80 KB, new limit 160KB/sec, refill period 1000 ms [COMPLETE] request size 80 KB, new limit 40KB/sec, refill period 1000 ms [COMPLETE] request size 160 KB, new limit 320KB/sec, refill period 1000 ms [COMPLETE] request size 160 KB, new limit 80KB/sec, refill period 1000 ms [ OK ] RateLimiterTest.LimitChangeTest (5002 ms) [----------] 4 tests from RateLimiterTest (27620 ms total) [----------] Global test environment tear-down [==========] 4 tests from 1 test case ran. (27621 ms total) [ PASSED ] 4 tests. ``` Reviewers: sdong, IslamAbdelRahman, yiwu, andrewkr Reviewed By: andrewkr Subscribers: andrewkr, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D60207
8 years ago
if (available_bytes_ < next_req->request_bytes) {
// Grant partial request_bytes to avoid starvation of requests
// that become asking for more bytes than available_bytes_
// due to dynamically reduced rate limiter's bytes_per_second that
// leads to reduced refill_bytes_per_period hence available_bytes_
fix rate limiter to avoid starvation Summary: The current implementation of rate limiter has the possibility to introduce resource starvation when change its limit. This diff aims to fix this problem by consuming request bytes partially. Test Plan: ``` ./rate_limiter_test [==========] Running 4 tests from 1 test case. [----------] Global test environment set-up. [----------] 4 tests from RateLimiterTest [ RUN ] RateLimiterTest.OverflowRate [ OK ] RateLimiterTest.OverflowRate (0 ms) [ RUN ] RateLimiterTest.StartStop [ OK ] RateLimiterTest.StartStop (0 ms) [ RUN ] RateLimiterTest.Rate request size [1 - 1023], limit 10 KB/sec, actual rate: 10.355712 KB/sec, elapsed 2.00 seconds request size [1 - 1023], limit 20 KB/sec, actual rate: 19.136564 KB/sec, elapsed 2.00 seconds request size [1 - 2047], limit 20 KB/sec, actual rate: 20.783976 KB/sec, elapsed 2.10 seconds request size [1 - 2047], limit 40 KB/sec, actual rate: 39.308144 KB/sec, elapsed 2.10 seconds request size [1 - 4095], limit 40 KB/sec, actual rate: 40.318349 KB/sec, elapsed 2.20 seconds request size [1 - 4095], limit 80 KB/sec, actual rate: 79.667396 KB/sec, elapsed 2.20 seconds request size [1 - 8191], limit 80 KB/sec, actual rate: 81.807158 KB/sec, elapsed 2.30 seconds request size [1 - 8191], limit 160 KB/sec, actual rate: 160.659761 KB/sec, elapsed 2.20 seconds request size [1 - 16383], limit 160 KB/sec, actual rate: 160.700990 KB/sec, elapsed 3.00 seconds request size [1 - 16383], limit 320 KB/sec, actual rate: 317.639481 KB/sec, elapsed 2.50 seconds [ OK ] RateLimiterTest.Rate (22618 ms) [ RUN ] RateLimiterTest.LimitChangeTest [COMPLETE] request size 10 KB, new limit 20KB/sec, refill period 1000 ms [COMPLETE] request size 10 KB, new limit 5KB/sec, refill period 1000 ms [COMPLETE] request size 20 KB, new limit 40KB/sec, refill period 1000 ms [COMPLETE] request size 20 KB, new limit 10KB/sec, refill period 1000 ms [COMPLETE] request size 40 KB, new limit 80KB/sec, refill period 1000 ms [COMPLETE] request size 40 KB, new limit 20KB/sec, refill period 1000 ms [COMPLETE] request size 80 KB, new limit 160KB/sec, refill period 1000 ms [COMPLETE] request size 80 KB, new limit 40KB/sec, refill period 1000 ms [COMPLETE] request size 160 KB, new limit 320KB/sec, refill period 1000 ms [COMPLETE] request size 160 KB, new limit 80KB/sec, refill period 1000 ms [ OK ] RateLimiterTest.LimitChangeTest (5002 ms) [----------] 4 tests from RateLimiterTest (27620 ms total) [----------] Global test environment tear-down [==========] 4 tests from 1 test case ran. (27621 ms total) [ PASSED ] 4 tests. ``` Reviewers: sdong, IslamAbdelRahman, yiwu, andrewkr Reviewed By: andrewkr Subscribers: andrewkr, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D60207
8 years ago
next_req->request_bytes -= available_bytes_;
available_bytes_ = 0;
break;
}
fix rate limiter to avoid starvation Summary: The current implementation of rate limiter has the possibility to introduce resource starvation when change its limit. This diff aims to fix this problem by consuming request bytes partially. Test Plan: ``` ./rate_limiter_test [==========] Running 4 tests from 1 test case. [----------] Global test environment set-up. [----------] 4 tests from RateLimiterTest [ RUN ] RateLimiterTest.OverflowRate [ OK ] RateLimiterTest.OverflowRate (0 ms) [ RUN ] RateLimiterTest.StartStop [ OK ] RateLimiterTest.StartStop (0 ms) [ RUN ] RateLimiterTest.Rate request size [1 - 1023], limit 10 KB/sec, actual rate: 10.355712 KB/sec, elapsed 2.00 seconds request size [1 - 1023], limit 20 KB/sec, actual rate: 19.136564 KB/sec, elapsed 2.00 seconds request size [1 - 2047], limit 20 KB/sec, actual rate: 20.783976 KB/sec, elapsed 2.10 seconds request size [1 - 2047], limit 40 KB/sec, actual rate: 39.308144 KB/sec, elapsed 2.10 seconds request size [1 - 4095], limit 40 KB/sec, actual rate: 40.318349 KB/sec, elapsed 2.20 seconds request size [1 - 4095], limit 80 KB/sec, actual rate: 79.667396 KB/sec, elapsed 2.20 seconds request size [1 - 8191], limit 80 KB/sec, actual rate: 81.807158 KB/sec, elapsed 2.30 seconds request size [1 - 8191], limit 160 KB/sec, actual rate: 160.659761 KB/sec, elapsed 2.20 seconds request size [1 - 16383], limit 160 KB/sec, actual rate: 160.700990 KB/sec, elapsed 3.00 seconds request size [1 - 16383], limit 320 KB/sec, actual rate: 317.639481 KB/sec, elapsed 2.50 seconds [ OK ] RateLimiterTest.Rate (22618 ms) [ RUN ] RateLimiterTest.LimitChangeTest [COMPLETE] request size 10 KB, new limit 20KB/sec, refill period 1000 ms [COMPLETE] request size 10 KB, new limit 5KB/sec, refill period 1000 ms [COMPLETE] request size 20 KB, new limit 40KB/sec, refill period 1000 ms [COMPLETE] request size 20 KB, new limit 10KB/sec, refill period 1000 ms [COMPLETE] request size 40 KB, new limit 80KB/sec, refill period 1000 ms [COMPLETE] request size 40 KB, new limit 20KB/sec, refill period 1000 ms [COMPLETE] request size 80 KB, new limit 160KB/sec, refill period 1000 ms [COMPLETE] request size 80 KB, new limit 40KB/sec, refill period 1000 ms [COMPLETE] request size 160 KB, new limit 320KB/sec, refill period 1000 ms [COMPLETE] request size 160 KB, new limit 80KB/sec, refill period 1000 ms [ OK ] RateLimiterTest.LimitChangeTest (5002 ms) [----------] 4 tests from RateLimiterTest (27620 ms total) [----------] Global test environment tear-down [==========] 4 tests from 1 test case ran. (27621 ms total) [ PASSED ] 4 tests. ``` Reviewers: sdong, IslamAbdelRahman, yiwu, andrewkr Reviewed By: andrewkr Subscribers: andrewkr, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D60207
8 years ago
available_bytes_ -= next_req->request_bytes;
next_req->request_bytes = 0;
Implement superior user & mid IO priority level in GenericRateLimiter (#8595) Summary: Context: An extra IO_USER priority in rate limiter allows users to optionally charge WAL writes / SST reads to rate limiter at this priority level, which then has higher priority than IO_HIGH and IO_LOW. With an extra IO_USER priority, it allows users to better specify the relative urgency/importance among different requests in rate limiter. As a consequence, IO resource management can better prioritize and limit resource based on user's need. The IO_USER is implemented as superior priority in GenericRateLimiter, in the sense that its request queue will always be iterated first without being constrained to fairness. The reason is that the notion of fairness is only meaningful in helping lower priorities in background IO (i.e, IO_HIGH/MID/LOW) to gain some fair chance to run so that it does not block foreground IO (i.e, the ones that are charged at the level of IO_USER). As we can see, the ultimate goal here is to not blocking foreground IO at IO_USER level, which justifies the superiority of IO_USER. Similar benefits exist for IO_MID priority. - Rewrote the logic of deciding the order of iterating request queues of high/low priorities to include the extra user/mid priority w/o affecting the existing behavior (see PR's [comment](https://github.com/facebook/rocksdb/pull/8595/files#r678749331)) - Included the request queue of user-pri/mid-pri in the code path of next-leader-candidate signaling and GenericRateLimiter's destructor - Included the extra user/mid-pri in bookkeeping data structures: total_bytes_through_ and total_requests_ - Re-written the previous impl of explicitly iterating priorities with a loop from Env::IO_LOW to Env::IO_TOTAL Pull Request resolved: https://github.com/facebook/rocksdb/pull/8595 Test Plan: - passed existing rate_limiter_test.cc - passed added unit tests in rate_limiter_test.cc - run performance test to verify performance with only high/low requests is not affected by this change - Set-up command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom --duration=5 --compression_type=none --num=100000000 --disable_auto_compactions=true --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1))` - Test command: `TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=overwrite --use_existing_db=true --disable_wal=true --duration=30 --compression_type=none --num=100000000 --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1)) --statistics=true --rate_limiter_bytes_per_sec=1048576 --rate_limiter_refill_period_us=1000 --threads=32 |& grep -E '(flush|compact)\.write\.bytes'` - Before (on branch upstream/master): `rocksdb.compact.write.bytes COUNT : 4014162` `rocksdb.flush.write.bytes COUNT : 26715832` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.66 - After (on branch rate_limiter_user_pri): `rocksdb.compact.write.bytes COUNT : 3807822` `rocksdb.flush.write.bytes COUNT : 26098659` rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.85 Reviewed By: ajkr Differential Revision: D30577783 Pulled By: hx235 fbshipit-source-id: 0881f2705ffd13ecd331256bde7e8ec874a353f4
3 years ago
total_bytes_through_[current_pri] += next_req->bytes;
queue->pop_front();
next_req->granted = true;
Simplify GenericRateLimiter algorithm (#8602) Summary: `GenericRateLimiter` slow path handles requests that cannot be satisfied immediately. Such requests enter a queue, and their thread stays in `Request()` until they are granted or the rate limiter is stopped. These threads are responsible for unblocking themselves. The work to do so is split into two main duties. (1) Waiting for the next refill time. (2) Refilling the bytes and granting requests. Prior to this PR, the slow path logic involved a leader election algorithm to pick one thread to perform (1) followed by (2). It elected the thread whose request was at the front of the highest priority non-empty queue since that request was most likely to be granted. This algorithm was efficient in terms of reducing intermediate wakeups, which is a thread waking up only to resume waiting after finding its request is not granted. However, the conceptual complexity of this algorithm was too high. It took me a long time to draw a timeline to understand how it works for just one edge case yet there were so many. This PR drops the leader election to reduce conceptual complexity. Now, the two duties can be performed by whichever thread acquires the lock first. The risk of this change is increasing the number of intermediate wakeups, however, we took steps to mitigate that. - `wait_until_refill_pending_` flag ensures only one thread performs (1). This\ prevents the thundering herd problem at the next refill time. The remaining\ threads wait on their condition variable with an unbounded duration -- thus we\ must remember to notify them to ensure forward progress. - (1) is typically done by a thread at the front of a queue. This is trivial\ when the queues are initially empty as the first choice that arrives must be\ the only entry in its queue. When queues are initially non-empty, we achieve\ this by having (2) notify a thread at the front of a queue (preferring higher\ priority) to perform the next duty. - We do not require any additional wakeup for (2). Typically it will just be\ done by the thread that finished (1). Combined, the second and third bullet points above suggest the refill/granting will typically be done by a request at the front of its queue. This is important because one wakeup is saved when a granted request happens to be in an already running thread. Note there are a few cases that still lead to intermediate wakeup, however. The first two are existing issues that also apply to the old algorithm, however, the third (including both subpoints) is new. - No request may be granted (only possible when rate limit dynamically\ decreases). - Requests from a different queue may be granted. - (2) may be run by a non-front request thread causing it to not be granted even\ if some requests in that same queue are granted. It can happen for a couple\ (unlikely) reasons. - A new request may sneak in and grab the lock at the refill time, before the\ thread finishing (1) can wake up and grab it. - A new request may sneak in and grab the lock and execute (1) before (2)'s\ chosen candidate can wake up and grab the lock. Then that non-front request\ thread performing (1) can carry over to perform (2). Pull Request resolved: https://github.com/facebook/rocksdb/pull/8602 Test Plan: - Use existing tests. The edge cases listed in the comment are all performance\ related; I could not really think of any related to correctness. The logic\ looks the same whether a thread wakes up/finishes its work early/on-time/late,\ or whether the thread is chosen vs. "steals" the work. - Verified write throughput and CPU overhead are basically the same with and\ without this change, even in a rate limiter heavy workload: Test command: ``` $ rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -benchmarks=fillrandom -num_multi_db=64 -num_low_pri_threads=64 -num_high_pri_threads=64 -write_buffer_size=262144 -target_file_size_base=262144 -max_bytes_for_level_base=1048576 -rate_limiter_bytes_per_sec=16777216 -key_size=24 -value_size=1000 -num=10000 -compression_type=none -rate_limiter_refill_period_us=1000 ``` Results before this PR: ``` fillrandom : 108.463 micros/op 9219 ops/sec; 9.0 MB/s 7.40user 8.84system 1:26.20elapsed 18%CPU (0avgtext+0avgdata 256140maxresident)k ``` Results after this PR: ``` fillrandom : 108.108 micros/op 9250 ops/sec; 9.0 MB/s 7.45user 8.23system 1:26.68elapsed 18%CPU (0avgtext+0avgdata 255688maxresident)k ``` Reviewed By: hx235 Differential Revision: D30048013 Pulled By: ajkr fbshipit-source-id: 6741bba9d9dfbccab359806d725105817fef818b
3 years ago
// Quota granted, signal the thread to exit
next_req->cv.Signal();
}
}
}
int64_t GenericRateLimiter::CalculateRefillBytesPerPeriod(
int64_t rate_bytes_per_sec) {
if (std::numeric_limits<int64_t>::max() / rate_bytes_per_sec <
options_.refill_period_us) {
// Avoid unexpected result in the overflow case. The result now is still
// inaccurate but is a number that is large enough.
return std::numeric_limits<int64_t>::max() / 1000000;
} else {
return rate_bytes_per_sec * options_.refill_period_us / 1000000;
}
}
Status GenericRateLimiter::Tune() {
const int kLowWatermarkPct = 50;
const int kHighWatermarkPct = 90;
const int kAdjustFactorPct = 5;
// computed rate limit will be in
// `[max_bytes_per_sec_ / kAllowedRangeFactor, max_bytes_per_sec_]`.
const int kAllowedRangeFactor = 20;
std::chrono::microseconds prev_tuned_time = tuned_time_;
tuned_time_ = std::chrono::microseconds(NowMicrosMonotonic());
int64_t elapsed_intervals =
(tuned_time_ - prev_tuned_time +
std::chrono::microseconds(options_.refill_period_us) -
std::chrono::microseconds(1)) /
std::chrono::microseconds(options_.refill_period_us);
// We tune every kRefillsPerTune intervals, so the overflow and division-by-
// zero conditions should never happen.
assert(num_drains_ <= std::numeric_limits<int64_t>::max() / 100);
assert(elapsed_intervals > 0);
int64_t drained_pct = num_drains_ * 100 / elapsed_intervals;
int64_t prev_bytes_per_sec = GetBytesPerSecond();
int64_t new_bytes_per_sec;
if (drained_pct == 0) {
new_bytes_per_sec = options_.max_bytes_per_sec / kAllowedRangeFactor;
} else if (drained_pct < kLowWatermarkPct) {
// sanitize to prevent overflow
int64_t sanitized_prev_bytes_per_sec =
std::min(prev_bytes_per_sec, std::numeric_limits<int64_t>::max() / 100);
new_bytes_per_sec =
std::max(options_.max_bytes_per_sec / kAllowedRangeFactor,
sanitized_prev_bytes_per_sec * 100 / (100 + kAdjustFactorPct));
} else if (drained_pct > kHighWatermarkPct) {
// sanitize to prevent overflow
int64_t sanitized_prev_bytes_per_sec =
std::min(prev_bytes_per_sec, std::numeric_limits<int64_t>::max() /
(100 + kAdjustFactorPct));
new_bytes_per_sec =
std::min(options_.max_bytes_per_sec,
sanitized_prev_bytes_per_sec * (100 + kAdjustFactorPct) / 100);
} else {
new_bytes_per_sec = prev_bytes_per_sec;
}
if (new_bytes_per_sec != prev_bytes_per_sec) {
SetBytesPerSecond(new_bytes_per_sec);
}
num_drains_ = 0;
return Status::OK();
}
RateLimiter* NewGenericRateLimiter(
int64_t rate_bytes_per_sec, int64_t refill_period_us /* = 100 * 1000 */,
int32_t fairness /* = 10 */,
RateLimiter::Mode mode /* = RateLimiter::Mode::kWritesOnly */,
bool auto_tuned /* = false */) {
assert(rate_bytes_per_sec > 0);
assert(refill_period_us > 0);
assert(fairness > 0);
std::unique_ptr<RateLimiter> limiter(
new GenericRateLimiter(rate_bytes_per_sec, refill_period_us, fairness,
mode, SystemClock::Default(), auto_tuned));
Status s = limiter->PrepareOptions(ConfigOptions());
if (s.ok()) {
return limiter.release();
} else {
assert(false);
return nullptr;
}
}
namespace {
#ifndef ROCKSDB_LITE
static int RegisterBuiltinRateLimiters(ObjectLibrary& library,
const std::string& /*arg*/) {
library.AddFactory<RateLimiter>(
GenericRateLimiter::kClassName(),
[](const std::string& /*uri*/, std::unique_ptr<RateLimiter>* guard,
std::string* /*errmsg*/) {
guard->reset(
new GenericRateLimiter(std::numeric_limits<int64_t>::max()));
return guard->get();
});
size_t num_types;
return static_cast<int>(library.GetFactoryCount(&num_types));
}
static std::unordered_map<std::string, RateLimiter::Mode>
rate_limiter_mode_map = {
{"kReadsOnly", RateLimiter::Mode::kReadsOnly},
{"kWritesOnly", RateLimiter::Mode::kWritesOnly},
{"kAllIo", RateLimiter::Mode::kAllIo},
};
#endif // ROCKSDB_LITE
static bool LoadRateLimiter(const std::string& name,
std::shared_ptr<RateLimiter>* limiter) {
auto plen = strlen(GenericRateLimiter::kClassName());
if (name.size() > plen + 2 && name[plen] == ':' &&
StartsWith(name, GenericRateLimiter::kClassName())) {
auto rate = ParseInt64(name.substr(plen + 1));
limiter->reset(new GenericRateLimiter(rate));
return true;
} else {
return false;
}
}
static std::unordered_map<std::string, OptionTypeInfo> rate_limiter_type_info =
{
#ifndef ROCKSDB_LITE
{"mode",
OptionTypeInfo::Enum<RateLimiter::Mode>(0, &rate_limiter_mode_map)},
#endif // ROCKSDB_LITE
};
} // namespace
RateLimiter::RateLimiter(Mode mode) : mode_(mode) {
RegisterOptions("", &mode_, &rate_limiter_type_info);
}
Status RateLimiter::CreateFromString(const ConfigOptions& config_options,
const std::string& value,
std::shared_ptr<RateLimiter>* result) {
if (value.empty()) {
result->reset();
return Status::OK();
} else {
#ifndef ROCKSDB_LITE
static std::once_flag once;
std::call_once(once, [&]() {
RegisterBuiltinRateLimiters(*(ObjectLibrary::Default().get()), "");
});
#endif // ROCKSDB_LITE
return LoadSharedObject<RateLimiter>(config_options, value, LoadRateLimiter,
result);
}
}
} // namespace ROCKSDB_NAMESPACE