rocksdb/monitoring/histogram_windowing.cc
mrambacher 12f1137355 Add a SystemClock class to capture the time functions of an Env (#7858)
Summary:
Introduces and uses a SystemClock class to RocksDB.  This class contains the time-related functions of an Env and these functions can be redirected from the Env to the SystemClock.

Many of the places that used an Env (Timer, PerfStepTimer, RepeatableThread, RateLimiter, WriteController) for time-related functions have been changed to use SystemClock instead.  There are likely more places that can be changed, but this is a start to show what can/should be done.  Over time it would be nice to migrate most (if not all) of the uses of the time functions from the Env to the SystemClock.

There are several Env classes that implement these functions.  Most of these have not been converted yet to SystemClock implementations; that will come in a subsequent PR.  It would be good to unify many of the Mock Timer implementations, so that they behave similarly and be tested similarly (some override Sleep, some use a MockSleep, etc).

Additionally, this change will allow new methods to be introduced to the SystemClock (like https://github.com/facebook/rocksdb/issues/7101 WaitFor) in a consistent manner across a smaller number of classes.

Pull Request resolved: https://github.com/facebook/rocksdb/pull/7858

Reviewed By: pdillinger

Differential Revision: D26006406

Pulled By: mrambacher

fbshipit-source-id: ed10a8abbdab7ff2e23d69d85bd25b3e7e899e90
2021-01-25 22:09:11 -08:00

203 lines
6.4 KiB
C++

// Copyright (c) 2013, 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 "monitoring/histogram_windowing.h"
#include <algorithm>
#include "monitoring/histogram.h"
#include "rocksdb/system_clock.h"
#include "util/cast_util.h"
namespace ROCKSDB_NAMESPACE {
HistogramWindowingImpl::HistogramWindowingImpl() {
clock_ = SystemClock::Default();
window_stats_.reset(new HistogramStat[static_cast<size_t>(num_windows_)]);
Clear();
}
HistogramWindowingImpl::HistogramWindowingImpl(
uint64_t num_windows,
uint64_t micros_per_window,
uint64_t min_num_per_window) :
num_windows_(num_windows),
micros_per_window_(micros_per_window),
min_num_per_window_(min_num_per_window) {
clock_ = SystemClock::Default();
window_stats_.reset(new HistogramStat[static_cast<size_t>(num_windows_)]);
Clear();
}
HistogramWindowingImpl::~HistogramWindowingImpl() {
}
void HistogramWindowingImpl::Clear() {
std::lock_guard<std::mutex> lock(mutex_);
stats_.Clear();
for (size_t i = 0; i < num_windows_; i++) {
window_stats_[i].Clear();
}
current_window_.store(0, std::memory_order_relaxed);
last_swap_time_.store(clock_->NowMicros(), std::memory_order_relaxed);
}
bool HistogramWindowingImpl::Empty() const { return stats_.Empty(); }
// This function is designed to be lock free, as it's in the critical path
// of any operation.
// Each individual value is atomic, it is just that some samples can go
// in the older bucket which is tolerable.
void HistogramWindowingImpl::Add(uint64_t value){
TimerTick();
// Parent (global) member update
stats_.Add(value);
// Current window update
window_stats_[static_cast<size_t>(current_window())].Add(value);
}
void HistogramWindowingImpl::Merge(const Histogram& other) {
if (strcmp(Name(), other.Name()) == 0) {
Merge(*static_cast_with_check<const HistogramWindowingImpl>(&other));
}
}
void HistogramWindowingImpl::Merge(const HistogramWindowingImpl& other) {
std::lock_guard<std::mutex> lock(mutex_);
stats_.Merge(other.stats_);
if (stats_.num_buckets_ != other.stats_.num_buckets_ ||
micros_per_window_ != other.micros_per_window_) {
return;
}
uint64_t cur_window = current_window();
uint64_t other_cur_window = other.current_window();
// going backwards for alignment
for (unsigned int i = 0;
i < std::min(num_windows_, other.num_windows_); i++) {
uint64_t window_index =
(cur_window + num_windows_ - i) % num_windows_;
uint64_t other_window_index =
(other_cur_window + other.num_windows_ - i) % other.num_windows_;
size_t windex = static_cast<size_t>(window_index);
size_t other_windex = static_cast<size_t>(other_window_index);
window_stats_[windex].Merge(
other.window_stats_[other_windex]);
}
}
std::string HistogramWindowingImpl::ToString() const {
return stats_.ToString();
}
double HistogramWindowingImpl::Median() const {
return Percentile(50.0);
}
double HistogramWindowingImpl::Percentile(double p) const {
// Retry 3 times in total
for (int retry = 0; retry < 3; retry++) {
uint64_t start_num = stats_.num();
double result = stats_.Percentile(p);
// Detect if swap buckets or Clear() was called during calculation
if (stats_.num() >= start_num) {
return result;
}
}
return 0.0;
}
double HistogramWindowingImpl::Average() const {
return stats_.Average();
}
double HistogramWindowingImpl::StandardDeviation() const {
return stats_.StandardDeviation();
}
void HistogramWindowingImpl::Data(HistogramData * const data) const {
stats_.Data(data);
}
void HistogramWindowingImpl::TimerTick() {
uint64_t curr_time = clock_->NowMicros();
size_t curr_window_ = static_cast<size_t>(current_window());
if (curr_time - last_swap_time() > micros_per_window_ &&
window_stats_[curr_window_].num() >= min_num_per_window_) {
SwapHistoryBucket();
}
}
void HistogramWindowingImpl::SwapHistoryBucket() {
// Threads executing Add() would be competing for this mutex, the first one
// who got the metex would take care of the bucket swap, other threads
// can skip this.
// If mutex is held by Merge() or Clear(), next Add() will take care of the
// swap, if needed.
if (mutex_.try_lock()) {
last_swap_time_.store(clock_->NowMicros(), std::memory_order_relaxed);
uint64_t curr_window = current_window();
uint64_t next_window = (curr_window == num_windows_ - 1) ?
0 : curr_window + 1;
// subtract next buckets from totals and swap to next buckets
HistogramStat& stats_to_drop =
window_stats_[static_cast<size_t>(next_window)];
if (!stats_to_drop.Empty()) {
for (size_t b = 0; b < stats_.num_buckets_; b++){
stats_.buckets_[b].fetch_sub(
stats_to_drop.bucket_at(b), std::memory_order_relaxed);
}
if (stats_.min() == stats_to_drop.min()) {
uint64_t new_min = std::numeric_limits<uint64_t>::max();
for (unsigned int i = 0; i < num_windows_; i++) {
if (i != next_window) {
uint64_t m = window_stats_[i].min();
if (m < new_min) new_min = m;
}
}
stats_.min_.store(new_min, std::memory_order_relaxed);
}
if (stats_.max() == stats_to_drop.max()) {
uint64_t new_max = 0;
for (unsigned int i = 0; i < num_windows_; i++) {
if (i != next_window) {
uint64_t m = window_stats_[i].max();
if (m > new_max) new_max = m;
}
}
stats_.max_.store(new_max, std::memory_order_relaxed);
}
stats_.num_.fetch_sub(stats_to_drop.num(), std::memory_order_relaxed);
stats_.sum_.fetch_sub(stats_to_drop.sum(), std::memory_order_relaxed);
stats_.sum_squares_.fetch_sub(
stats_to_drop.sum_squares(), std::memory_order_relaxed);
stats_to_drop.Clear();
}
// advance to next window bucket
current_window_.store(next_window, std::memory_order_relaxed);
mutex_.unlock();
}
}
} // namespace ROCKSDB_NAMESPACE