Summary: A generic rate limiter that can be shared by threads and rocksdb instances. Will use this to smooth out write traffic generated by compaction and flush. This will help us get better p99 behavior on flash storage. Test Plan: unit test output ==== Test RateLimiterTest.Rate request size [1 - 1023], limit 10 KB/sec, actual rate: 10.374969 KB/sec, elapsed 2002265 request size [1 - 2047], limit 20 KB/sec, actual rate: 20.771242 KB/sec, elapsed 2002139 request size [1 - 4095], limit 40 KB/sec, actual rate: 41.285299 KB/sec, elapsed 2202424 request size [1 - 8191], limit 80 KB/sec, actual rate: 81.371605 KB/sec, elapsed 2402558 request size [1 - 16383], limit 160 KB/sec, actual rate: 162.541268 KB/sec, elapsed 3303500 Reviewers: yhchiang, igor, sdong Reviewed By: sdong Subscribers: leveldb Differential Revision: https://reviews.facebook.net/D19359main
Lei Jin
11 years ago
parent
b278ae8e50
commit
5ef1ba7ff5
@ -0,0 +1,200 @@ |
||||
// Copyright (c) 2014, Facebook, Inc. All rights reserved.
|
||||
// This source code is licensed under the BSD-style license found in the
|
||||
// LICENSE file in the root directory of this source tree. An additional grant
|
||||
// of patent rights can be found in the PATENTS file in the same 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 "rocksdb/env.h" |
||||
|
||||
namespace rocksdb { |
||||
|
||||
|
||||
// Pending request
|
||||
struct RateLimiter::Req { |
||||
explicit Req(int64_t bytes, port::Mutex* mu) : |
||||
bytes(bytes), cv(mu), granted(false) {} |
||||
int64_t bytes; |
||||
port::CondVar cv; |
||||
bool granted; |
||||
}; |
||||
|
||||
|
||||
RateLimiter::RateLimiter(int64_t rate_bytes_per_sec, int64_t refill_period_us, |
||||
int32_t fairness) |
||||
: refill_period_us_(refill_period_us), |
||||
refill_bytes_per_period_(rate_bytes_per_sec * refill_period_us / 1000000.0), |
||||
env_(Env::Default()), |
||||
stop_(false), |
||||
exit_cv_(&request_mutex_), |
||||
requests_to_wait_(0), |
||||
total_requests_{0, 0}, |
||||
total_bytes_through_{0, 0}, |
||||
available_bytes_(0), |
||||
next_refill_us_(env_->NowMicros()), |
||||
fairness_(fairness > 100 ? 100 : fairness), |
||||
rnd_((uint32_t)time(nullptr)), |
||||
leader_(nullptr) { |
||||
total_bytes_through_[0] = 0; |
||||
total_bytes_through_[1] = 0; |
||||
} |
||||
|
||||
RateLimiter::~RateLimiter() { |
||||
MutexLock g(&request_mutex_); |
||||
stop_ = true; |
||||
requests_to_wait_ = queue_[Env::IO_LOW].size() + queue_[Env::IO_HIGH].size(); |
||||
for (auto& r : queue_[Env::IO_HIGH]) { |
||||
r->cv.Signal(); |
||||
} |
||||
for (auto& r : queue_[Env::IO_LOW]) { |
||||
r->cv.Signal(); |
||||
} |
||||
while (requests_to_wait_ > 0) { |
||||
exit_cv_.Wait(); |
||||
} |
||||
} |
||||
|
||||
void RateLimiter::Request(int64_t bytes, const Env::IOPriority pri) { |
||||
assert(bytes < refill_bytes_per_period_); |
||||
|
||||
MutexLock g(&request_mutex_); |
||||
if (stop_) { |
||||
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); |
||||
|
||||
do { |
||||
bool timedout = false; |
||||
// Leader election, candidates can be:
|
||||
// (1) a new incoming request,
|
||||
// (2) a previous leader, whose quota has not been not assigned yet due
|
||||
// to lower priority
|
||||
// (3) a previous waiter at the front of queue, who got notified by
|
||||
// previous leader
|
||||
if (leader_ == nullptr && |
||||
((!queue_[Env::IO_HIGH].empty() && |
||||
&r == queue_[Env::IO_HIGH].front()) || |
||||
(!queue_[Env::IO_LOW].empty() && |
||||
&r == queue_[Env::IO_LOW].front()))) { |
||||
leader_ = &r; |
||||
timedout = r.cv.TimedWait(next_refill_us_); |
||||
} else { |
||||
// Not at the front of queue or an leader has already been elected
|
||||
r.cv.Wait(); |
||||
} |
||||
|
||||
// request_mutex_ is held from now on
|
||||
if (stop_) { |
||||
--requests_to_wait_; |
||||
exit_cv_.Signal(); |
||||
return; |
||||
} |
||||
|
||||
// Make sure the waken up request is always the header of its queue
|
||||
assert(r.granted || |
||||
(!queue_[Env::IO_HIGH].empty() && |
||||
&r == queue_[Env::IO_HIGH].front()) || |
||||
(!queue_[Env::IO_LOW].empty() && |
||||
&r == queue_[Env::IO_LOW].front())); |
||||
assert(leader_ == nullptr || |
||||
(!queue_[Env::IO_HIGH].empty() && |
||||
leader_ == queue_[Env::IO_HIGH].front()) || |
||||
(!queue_[Env::IO_LOW].empty() && |
||||
leader_ == queue_[Env::IO_LOW].front())); |
||||
|
||||
if (leader_ == &r) { |
||||
// Waken up from TimedWait()
|
||||
if (timedout) { |
||||
// Time to do refill!
|
||||
Refill(); |
||||
|
||||
// Re-elect a new leader regardless. This is to simplify the
|
||||
// election handling.
|
||||
leader_ = nullptr; |
||||
|
||||
// Notify the header of queue if current leader is going away
|
||||
if (r.granted) { |
||||
// Current leader already got granted with quota. Notify header
|
||||
// of waiting queue to participate next round of election.
|
||||
assert((queue_[Env::IO_HIGH].empty() || |
||||
&r != queue_[Env::IO_HIGH].front()) && |
||||
(queue_[Env::IO_LOW].empty() || |
||||
&r != queue_[Env::IO_LOW].front())); |
||||
if (!queue_[Env::IO_HIGH].empty()) { |
||||
queue_[Env::IO_HIGH].front()->cv.Signal(); |
||||
} else if (!queue_[Env::IO_LOW].empty()) { |
||||
queue_[Env::IO_LOW].front()->cv.Signal(); |
||||
} |
||||
// Done
|
||||
break; |
||||
} |
||||
} else { |
||||
// Spontaneous wake up, need to continue to wait
|
||||
assert(!r.granted); |
||||
leader_ = nullptr; |
||||
} |
||||
} else { |
||||
// Waken up by previous leader:
|
||||
// (1) if requested quota is granted, it is done.
|
||||
// (2) if requested quota is not granted, this means current thread
|
||||
// was picked as a new leader candidate (previous leader got quota).
|
||||
// It needs to participate leader election because a new request may
|
||||
// come in before this thread gets waken up. So it may actually need
|
||||
// to do Wait() again.
|
||||
assert(!timedout); |
||||
} |
||||
} while (!r.granted); |
||||
} |
||||
|
||||
void RateLimiter::Refill() { |
||||
next_refill_us_ = env_->NowMicros() + refill_period_us_; |
||||
// Carry over the left over quota from the last period
|
||||
if (available_bytes_ < refill_bytes_per_period_) { |
||||
available_bytes_ += refill_bytes_per_period_; |
||||
} |
||||
|
||||
int use_low_pri_first = rnd_.OneIn(fairness_) ? 0 : 1; |
||||
for (int q = 0; q < 2; ++q) { |
||||
auto use_pri = (use_low_pri_first == q) ? Env::IO_LOW : Env::IO_HIGH; |
||||
auto* queue = &queue_[use_pri]; |
||||
while (!queue->empty()) { |
||||
auto* next_req = queue->front(); |
||||
if (available_bytes_ < next_req->bytes) { |
||||
break; |
||||
} |
||||
available_bytes_ -= next_req->bytes; |
||||
total_bytes_through_[use_pri] += next_req->bytes; |
||||
queue->pop_front(); |
||||
|
||||
next_req->granted = true; |
||||
if (next_req != leader_) { |
||||
// Quota granted, signal the thread
|
||||
next_req->cv.Signal(); |
||||
} |
||||
} |
||||
} |
||||
} |
||||
|
||||
RateLimiter* NewRateLimiter( |
||||
int64_t rate_bytes_per_sec, int64_t refill_period_us, int32_t fairness) { |
||||
return new RateLimiter(rate_bytes_per_sec, refill_period_us, fairness); |
||||
} |
||||
|
||||
} // namespace rocksdb
|
||||
@ -0,0 +1,88 @@ |
||||
// Copyright (c) 2014, Facebook, Inc. All rights reserved.
|
||||
// This source code is licensed under the BSD-style license found in the
|
||||
// LICENSE file in the root directory of this source tree. An additional grant
|
||||
// of patent rights can be found in the PATENTS file in the same 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.
|
||||
|
||||
#pragma once |
||||
|
||||
#include <thread> |
||||
#include <deque> |
||||
|
||||
#include "port/port_posix.h" |
||||
#include "util/mutexlock.h" |
||||
#include "util/random.h" |
||||
#include "rocksdb/env.h" |
||||
|
||||
namespace rocksdb { |
||||
|
||||
class RateLimiter { |
||||
public: |
||||
RateLimiter(int64_t refill_bytes, int64_t refill_period_us, int32_t fairness); |
||||
|
||||
~RateLimiter(); |
||||
|
||||
// Request for token to write bytes. If this request can not be satisfied,
|
||||
// the call is blocked. Caller is responsible to make sure
|
||||
// bytes < GetSingleBurstBytes()
|
||||
void Request(const int64_t bytes, const Env::IOPriority pri); |
||||
|
||||
int64_t GetTotalBytesThrough( |
||||
const Env::IOPriority pri = Env::IO_TOTAL) const { |
||||
MutexLock g(&request_mutex_); |
||||
if (pri == Env::IO_TOTAL) { |
||||
return total_bytes_through_[Env::IO_LOW] + |
||||
total_bytes_through_[Env::IO_HIGH]; |
||||
} |
||||
return total_bytes_through_[pri]; |
||||
} |
||||
|
||||
int64_t GetTotalRequests(const Env::IOPriority pri = Env::IO_TOTAL) const { |
||||
MutexLock g(&request_mutex_); |
||||
if (pri == Env::IO_TOTAL) { |
||||
return total_requests_[Env::IO_LOW] + total_requests_[Env::IO_HIGH]; |
||||
} |
||||
return total_requests_[pri]; |
||||
} |
||||
|
||||
int64_t GetSingleBurstBytes() const { |
||||
// const var
|
||||
return refill_bytes_per_period_; |
||||
} |
||||
|
||||
int64_t GetAvailableBytes() const { |
||||
MutexLock g(&request_mutex_); |
||||
return available_bytes_; |
||||
} |
||||
|
||||
private: |
||||
void Refill(); |
||||
|
||||
// This mutex guard all internal states
|
||||
mutable port::Mutex request_mutex_; |
||||
|
||||
const int64_t refill_period_us_; |
||||
const int64_t refill_bytes_per_period_; |
||||
Env* const env_; |
||||
|
||||
bool stop_; |
||||
port::CondVar exit_cv_; |
||||
int32_t requests_to_wait_; |
||||
|
||||
int64_t total_requests_[Env::IO_TOTAL]; |
||||
int64_t total_bytes_through_[Env::IO_TOTAL]; |
||||
int64_t available_bytes_; |
||||
int64_t next_refill_us_; |
||||
|
||||
int32_t fairness_; |
||||
Random rnd_; |
||||
|
||||
struct Req; |
||||
Req* leader_; |
||||
std::deque<Req*> queue_[Env::IO_TOTAL]; |
||||
}; |
||||
|
||||
} // namespace rocksdb
|
||||
@ -0,0 +1,80 @@ |
||||
// Copyright (c) 2014, Facebook, Inc. All rights reserved.
|
||||
// This source code is licensed under the BSD-style license found in the
|
||||
// LICENSE file in the root directory of this source tree. An additional grant
|
||||
// of patent rights can be found in the PATENTS file in the same 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 <limits> |
||||
#include "util/testharness.h" |
||||
#include "util/rate_limiter.h" |
||||
#include "util/random.h" |
||||
#include "rocksdb/env.h" |
||||
|
||||
namespace rocksdb { |
||||
|
||||
class RateLimiterTest { |
||||
}; |
||||
|
||||
TEST(RateLimiterTest, StartStop) { |
||||
std::unique_ptr<RateLimiter> limiter(new RateLimiter(100, 100, 10)); |
||||
} |
||||
|
||||
TEST(RateLimiterTest, Rate) { |
||||
auto* env = Env::Default(); |
||||
struct Arg { |
||||
Arg(int64_t target_rate, int burst) |
||||
: limiter(new RateLimiter(target_rate, 100 * 1000, 10)), |
||||
request_size(target_rate / 10), |
||||
burst(burst) {} |
||||
std::unique_ptr<RateLimiter> limiter; |
||||
int64_t request_size; |
||||
int burst; |
||||
}; |
||||
|
||||
auto writer = [](void* p) { |
||||
auto* env = Env::Default(); |
||||
auto* arg = static_cast<Arg*>(p); |
||||
// Test for 2 seconds
|
||||
auto until = env->NowMicros() + 2 * 1000000; |
||||
Random r((uint32_t)(env->NowNanos() % |
||||
std::numeric_limits<uint32_t>::max())); |
||||
while (env->NowMicros() < until) { |
||||
for (int i = 0; i < static_cast<int>(r.Skewed(arg->burst) + 1); ++i) { |
||||
arg->limiter->Request(r.Uniform(arg->request_size - 1) + 1, |
||||
Env::IO_HIGH); |
||||
} |
||||
arg->limiter->Request(r.Uniform(arg->request_size - 1) + 1, |
||||
Env::IO_LOW); |
||||
} |
||||
}; |
||||
|
||||
for (int i = 1; i <= 16; i*=2) { |
||||
int64_t target = i * 1024 * 10; |
||||
Arg arg(target, i / 4 + 1); |
||||
auto start = env->NowMicros(); |
||||
for (int t = 0; t < i; ++t) { |
||||
env->StartThread(writer, &arg); |
||||
} |
||||
env->WaitForJoin(); |
||||
|
||||
auto elapsed = env->NowMicros() - start; |
||||
double rate = arg.limiter->GetTotalBytesThrough() |
||||
* 1000000.0 / elapsed; |
||||
fprintf(stderr, "request size [1 - %ld], limit %ld KB/sec, " |
||||
"actual rate: %lf KB/sec, elapsed %.2lf seconds\n", |
||||
arg.request_size - 1, target / 1024, rate / 1024, |
||||
elapsed / 1000000.0); |
||||
|
||||
ASSERT_GE(rate / target, 0.95); |
||||
ASSERT_LE(rate / target, 1.05); |
||||
} |
||||
} |
||||
|
||||
} // namespace rocksdb
|
||||
|
||||
int main(int argc, char** argv) { |
||||
return rocksdb::test::RunAllTests(); |
||||
} |
||||
Loading…
Reference in new issue