|
|
|
// Copyright (c) 2011-present, 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.
|
|
|
|
// This source code is also licensed under the GPLv2 license found in the
|
|
|
|
// COPYING file in the root directory of this source tree.
|
|
|
|
//
|
|
|
|
#pragma once
|
|
|
|
#include "rocksdb/statistics.h"
|
|
|
|
|
|
|
|
#include <vector>
|
|
|
|
#include <atomic>
|
|
|
|
#include <string>
|
|
|
|
|
|
|
|
#include "monitoring/histogram.h"
|
|
|
|
#include "port/likely.h"
|
|
|
|
#include "port/port.h"
|
|
|
|
#include "util/mutexlock.h"
|
|
|
|
#include "util/thread_local.h"
|
|
|
|
|
|
|
|
namespace rocksdb {
|
|
|
|
|
|
|
|
enum TickersInternal : uint32_t {
|
|
|
|
INTERNAL_TICKER_ENUM_START = TICKER_ENUM_MAX,
|
|
|
|
INTERNAL_TICKER_ENUM_MAX
|
|
|
|
};
|
|
|
|
|
|
|
|
enum HistogramsInternal : uint32_t {
|
|
|
|
INTERNAL_HISTOGRAM_START = HISTOGRAM_ENUM_MAX,
|
|
|
|
INTERNAL_HISTOGRAM_ENUM_MAX
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
|
|
class StatisticsImpl : public Statistics {
|
|
|
|
public:
|
|
|
|
StatisticsImpl(std::shared_ptr<Statistics> stats,
|
|
|
|
bool enable_internal_stats);
|
|
|
|
virtual ~StatisticsImpl();
|
|
|
|
|
|
|
|
virtual uint64_t getTickerCount(uint32_t ticker_type) const override;
|
|
|
|
virtual void histogramData(uint32_t histogram_type,
|
|
|
|
HistogramData* const data) const override;
|
Add Statistics.getHistogramString() to print more detailed outputs of a histogram
Summary:
Provide a way for users to know more detailed ditribution of a histogram metrics. Example outputs:
Manually add statement
fprintf(stdout, "%s\n", dbstats->getHistogramString(SST_READ_MICROS).c_str());
Will print out something like:
Count: 989151 Average: 1.7659 StdDev: 1.52
Min: 0.0000 Median: 1.2071 Max: 860.0000
Percentiles: P50: 1.21 P75: 1.70 P99: 5.12 P99.9: 13.67 P99.99: 21.70
------------------------------------------------------
[ 0, 1 ) 390839 39.513% 39.513% ########
[ 1, 2 ) 500918 50.641% 90.154% ##########
[ 2, 3 ) 79358 8.023% 98.177% ##
[ 3, 4 ) 6297 0.637% 98.813%
[ 4, 5 ) 1712 0.173% 98.986%
[ 5, 6 ) 1134 0.115% 99.101%
[ 6, 7 ) 1222 0.124% 99.224%
[ 7, 8 ) 1529 0.155% 99.379%
[ 8, 9 ) 1264 0.128% 99.507%
[ 9, 10 ) 988 0.100% 99.607%
[ 10, 12 ) 1378 0.139% 99.746%
[ 12, 14 ) 1828 0.185% 99.931%
[ 14, 16 ) 410 0.041% 99.972%
[ 16, 18 ) 72 0.007% 99.980%
[ 18, 20 ) 67 0.007% 99.986%
[ 20, 25 ) 106 0.011% 99.997%
[ 25, 30 ) 24 0.002% 99.999%
[ 30, 35 ) 1 0.000% 100.000%
[ 250, 300 ) 2 0.000% 100.000%
[ 300, 350 ) 1 0.000% 100.000%
[ 800, 900 ) 1 0.000% 100.000%
Test Plan: Manually add a print in db_bench and make sure it prints out as expected. Will add some codes to cover the function
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D43611
9 years ago
|
|
|
std::string getHistogramString(uint32_t histogram_type) const override;
|
|
|
|
|
|
|
|
virtual void setTickerCount(uint32_t ticker_type, uint64_t count) override;
|
|
|
|
virtual uint64_t getAndResetTickerCount(uint32_t ticker_type) override;
|
|
|
|
virtual void recordTick(uint32_t ticker_type, uint64_t count) override;
|
|
|
|
virtual void measureTime(uint32_t histogram_type, uint64_t value) override;
|
|
|
|
|
|
|
|
virtual Status Reset() override;
|
|
|
|
virtual std::string ToString() const override;
|
|
|
|
virtual bool HistEnabledForType(uint32_t type) const override;
|
|
|
|
|
|
|
|
private:
|
|
|
|
std::shared_ptr<Statistics> stats_shared_;
|
|
|
|
Statistics* stats_;
|
|
|
|
bool enable_internal_stats_;
|
|
|
|
// Synchronizes anything that operates on other threads' thread-specific data
|
|
|
|
// such that operations like Reset() can be performed atomically.
|
|
|
|
mutable port::Mutex aggregate_lock_;
|
|
|
|
|
|
|
|
// Holds data maintained by each thread for implementing tickers.
|
|
|
|
struct ThreadTickerInfo {
|
|
|
|
std::atomic_uint_fast64_t value;
|
|
|
|
// During teardown, value will be summed into *merged_sum.
|
|
|
|
std::atomic_uint_fast64_t* merged_sum;
|
|
|
|
|
|
|
|
ThreadTickerInfo(uint_fast64_t _value,
|
|
|
|
std::atomic_uint_fast64_t* _merged_sum)
|
|
|
|
: value(_value), merged_sum(_merged_sum) {}
|
|
|
|
};
|
|
|
|
|
Thread-specific histogram statistics
Summary:
To reduce contention for atomics when HistogramStats are shared across
threads, this diff makes them thread-specific so updates are faster. This comes
at the expense of slower reads (much less frequent), which now require merging
all histograms. In this diff,
- Thread-specific HistogramImpl is created upon the thread's first measureTime()
- Thread-specific HistogramImpl are merged and deleted upon thread termination or ThreadLocalPtr destruction, whichever comes first
- getHistogramString() and histogramData() merge all histograms, both thread-specific and previously merged ones
Test Plan:
unit tests, ran db_bench and verified histograms look similar
before:
$ TEST_TMPDIR=/dev/shm/ perf record -g ./db_bench --benchmarks=readwhilewriting --statistics --num=1000000 --use_existing_db --threads=64 --cache_size=250000000 --compression_type=lz4
...
+ 7.63% db_bench db_bench [.] rocksdb::HistogramStat::Add
after:
$ TEST_TMPDIR=/dev/shm/ perf record -g ./db_bench --benchmarks=readwhilewriting --statistics --num=1000000 --use_existing_db --threads=64 --cache_size=250000000 --compression_type=lz4
...
+ 0.98% db_bench db_bench [.] rocksdb::HistogramStat::Add
Reviewers: sdong, MarkCallaghan, kradhakrishnan, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: andrewkr, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D62649
8 years ago
|
|
|
// Holds data maintained by each thread for implementing histograms.
|
|
|
|
struct ThreadHistogramInfo {
|
|
|
|
HistogramImpl value;
|
|
|
|
// During teardown, value will be merged into *merged_hist while holding
|
|
|
|
// *merge_lock, which also syncs with the merges necessary for reads.
|
|
|
|
HistogramImpl* merged_hist;
|
|
|
|
port::Mutex* merge_lock;
|
|
|
|
|
|
|
|
ThreadHistogramInfo(HistogramImpl* _merged_hist, port::Mutex* _merge_lock)
|
|
|
|
: value(), merged_hist(_merged_hist), merge_lock(_merge_lock) {}
|
|
|
|
};
|
|
|
|
|
|
|
|
// Holds global data for implementing tickers.
|
|
|
|
struct TickerInfo {
|
|
|
|
TickerInfo()
|
|
|
|
: thread_value(new ThreadLocalPtr(&mergeThreadValue)), merged_sum(0) {}
|
|
|
|
// Holds thread-specific pointer to ThreadTickerInfo
|
|
|
|
std::unique_ptr<ThreadLocalPtr> thread_value;
|
|
|
|
// Sum of thread-specific values for tickers that have been reset due to
|
|
|
|
// thread termination or ThreadLocalPtr destruction. Also, this is used by
|
|
|
|
// setTickerCount() to conveniently change the global value by setting this
|
|
|
|
// while simultaneously zeroing all thread-local values.
|
|
|
|
std::atomic_uint_fast64_t merged_sum;
|
|
|
|
|
|
|
|
static void mergeThreadValue(void* ptr) {
|
|
|
|
auto info_ptr = static_cast<ThreadTickerInfo*>(ptr);
|
|
|
|
*info_ptr->merged_sum += info_ptr->value;
|
|
|
|
delete info_ptr;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
Thread-specific histogram statistics
Summary:
To reduce contention for atomics when HistogramStats are shared across
threads, this diff makes them thread-specific so updates are faster. This comes
at the expense of slower reads (much less frequent), which now require merging
all histograms. In this diff,
- Thread-specific HistogramImpl is created upon the thread's first measureTime()
- Thread-specific HistogramImpl are merged and deleted upon thread termination or ThreadLocalPtr destruction, whichever comes first
- getHistogramString() and histogramData() merge all histograms, both thread-specific and previously merged ones
Test Plan:
unit tests, ran db_bench and verified histograms look similar
before:
$ TEST_TMPDIR=/dev/shm/ perf record -g ./db_bench --benchmarks=readwhilewriting --statistics --num=1000000 --use_existing_db --threads=64 --cache_size=250000000 --compression_type=lz4
...
+ 7.63% db_bench db_bench [.] rocksdb::HistogramStat::Add
after:
$ TEST_TMPDIR=/dev/shm/ perf record -g ./db_bench --benchmarks=readwhilewriting --statistics --num=1000000 --use_existing_db --threads=64 --cache_size=250000000 --compression_type=lz4
...
+ 0.98% db_bench db_bench [.] rocksdb::HistogramStat::Add
Reviewers: sdong, MarkCallaghan, kradhakrishnan, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: andrewkr, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D62649
8 years ago
|
|
|
// Holds global data for implementing histograms.
|
|
|
|
struct HistogramInfo {
|
|
|
|
HistogramInfo()
|
|
|
|
: merged_hist(),
|
|
|
|
merge_lock(),
|
|
|
|
thread_value(new ThreadLocalPtr(&mergeThreadValue)) {}
|
|
|
|
// Merged thread-specific values for histograms that have been reset due to
|
|
|
|
// thread termination or ThreadLocalPtr destruction. Note these must be
|
|
|
|
// destroyed after thread_value since its destructor accesses them.
|
|
|
|
HistogramImpl merged_hist;
|
|
|
|
mutable port::Mutex merge_lock;
|
|
|
|
// Holds thread-specific pointer to ThreadHistogramInfo
|
|
|
|
std::unique_ptr<ThreadLocalPtr> thread_value;
|
|
|
|
|
|
|
|
static void mergeThreadValue(void* ptr) {
|
|
|
|
auto info_ptr = static_cast<ThreadHistogramInfo*>(ptr);
|
|
|
|
{
|
|
|
|
MutexLock lock(info_ptr->merge_lock);
|
|
|
|
info_ptr->merged_hist->Merge(info_ptr->value);
|
|
|
|
}
|
|
|
|
delete info_ptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Returns a histogram that merges all histograms (thread-specific and
|
|
|
|
// previously merged ones).
|
|
|
|
std::unique_ptr<HistogramImpl> getMergedHistogram() const;
|
|
|
|
};
|
|
|
|
|
|
|
|
uint64_t getTickerCountLocked(uint32_t ticker_type) const;
|
|
|
|
void histogramDataLocked(uint32_t histogram_type,
|
|
|
|
HistogramData* const data) const;
|
|
|
|
void setTickerCountLocked(uint32_t ticker_type, uint64_t count);
|
|
|
|
|
|
|
|
// Returns the info for this tickerType/thread. It sets a new info with zeroed
|
|
|
|
// counter if none exists.
|
Thread-specific histogram statistics
Summary:
To reduce contention for atomics when HistogramStats are shared across
threads, this diff makes them thread-specific so updates are faster. This comes
at the expense of slower reads (much less frequent), which now require merging
all histograms. In this diff,
- Thread-specific HistogramImpl is created upon the thread's first measureTime()
- Thread-specific HistogramImpl are merged and deleted upon thread termination or ThreadLocalPtr destruction, whichever comes first
- getHistogramString() and histogramData() merge all histograms, both thread-specific and previously merged ones
Test Plan:
unit tests, ran db_bench and verified histograms look similar
before:
$ TEST_TMPDIR=/dev/shm/ perf record -g ./db_bench --benchmarks=readwhilewriting --statistics --num=1000000 --use_existing_db --threads=64 --cache_size=250000000 --compression_type=lz4
...
+ 7.63% db_bench db_bench [.] rocksdb::HistogramStat::Add
after:
$ TEST_TMPDIR=/dev/shm/ perf record -g ./db_bench --benchmarks=readwhilewriting --statistics --num=1000000 --use_existing_db --threads=64 --cache_size=250000000 --compression_type=lz4
...
+ 0.98% db_bench db_bench [.] rocksdb::HistogramStat::Add
Reviewers: sdong, MarkCallaghan, kradhakrishnan, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: andrewkr, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D62649
8 years ago
|
|
|
ThreadTickerInfo* getThreadTickerInfo(uint32_t ticker_type);
|
|
|
|
// Returns the info for this histogramType/thread. It sets a new histogram
|
|
|
|
// with zeroed data if none exists.
|
|
|
|
ThreadHistogramInfo* getThreadHistogramInfo(uint32_t histogram_type);
|
|
|
|
|
Thread-specific histogram statistics
Summary:
To reduce contention for atomics when HistogramStats are shared across
threads, this diff makes them thread-specific so updates are faster. This comes
at the expense of slower reads (much less frequent), which now require merging
all histograms. In this diff,
- Thread-specific HistogramImpl is created upon the thread's first measureTime()
- Thread-specific HistogramImpl are merged and deleted upon thread termination or ThreadLocalPtr destruction, whichever comes first
- getHistogramString() and histogramData() merge all histograms, both thread-specific and previously merged ones
Test Plan:
unit tests, ran db_bench and verified histograms look similar
before:
$ TEST_TMPDIR=/dev/shm/ perf record -g ./db_bench --benchmarks=readwhilewriting --statistics --num=1000000 --use_existing_db --threads=64 --cache_size=250000000 --compression_type=lz4
...
+ 7.63% db_bench db_bench [.] rocksdb::HistogramStat::Add
after:
$ TEST_TMPDIR=/dev/shm/ perf record -g ./db_bench --benchmarks=readwhilewriting --statistics --num=1000000 --use_existing_db --threads=64 --cache_size=250000000 --compression_type=lz4
...
+ 0.98% db_bench db_bench [.] rocksdb::HistogramStat::Add
Reviewers: sdong, MarkCallaghan, kradhakrishnan, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: andrewkr, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D62649
8 years ago
|
|
|
TickerInfo tickers_[INTERNAL_TICKER_ENUM_MAX];
|
|
|
|
HistogramInfo histograms_[INTERNAL_HISTOGRAM_ENUM_MAX];
|
|
|
|
};
|
|
|
|
|
|
|
|
// Utility functions
|
|
|
|
inline void MeasureTime(Statistics* statistics, uint32_t histogram_type,
|
|
|
|
uint64_t value) {
|
|
|
|
if (statistics) {
|
|
|
|
statistics->measureTime(histogram_type, value);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
inline void RecordTick(Statistics* statistics, uint32_t ticker_type,
|
|
|
|
uint64_t count = 1) {
|
|
|
|
if (statistics) {
|
|
|
|
statistics->recordTick(ticker_type, count);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
inline void SetTickerCount(Statistics* statistics, uint32_t ticker_type,
|
|
|
|
uint64_t count) {
|
|
|
|
if (statistics) {
|
|
|
|
statistics->setTickerCount(ticker_type, count);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
}
|