提交 678ba5e4 编写于 作者: M matthewvon 提交者: Facebook GitHub Bot

SyncPoint::Process thrashes heap ... fix it (#9023)

Summary:
The first parameter of SyncPoint::Process is "const std::string&".  The majority, maybe all, of the actual calls to this function use a "const char *".  The conversion before entering the function requires a construction of a std::string object on the heap.  This std::object is then typically not needed because first use of the string is a rocksdb::Slice which has a less costly conversion of char * to slice.

Example:

We have a load and iterate test.  The test loads 10m keys and iterates most via 10 rocksdb::Iterator objects.  We used TCMALLOC to gather information about allocation and space usage during iterators.

- Before this PR:  test took 32 min 17 sec
- After this PR:  test took 1 min 14 sec

The TCMALLOC top object list before this PR:

<pre>
Total: 5105999 objects
 5003717  98.0%  98.0%  5009471  98.1% rocksdb::DBIter::MergeValuesNewToOld (inline)
   20260   0.4%  98.4%    20260   0.4% std::__cxx11::basic_string::_M_mutate
   15214   0.3%  98.7%    15214   0.3% rocksdb::UncompressBlockContentsForCompressionType (inline)
   13408   0.3%  99.0%    13408   0.3% std::_Rb_tree::_M_emplace_hint_unique [clone .constprop.416] (inline)
   12957   0.3%  99.2%    12957   0.3% std::_Rb_tree::_M_emplace_hint_unique [clone .constprop.405] (inline)
    9327   0.2%  99.4%     9327   0.2% std::_Rb_tree::_M_copy (inline)
    7691   0.2%  99.5%     9919   0.2% JVM_FindSignal
    2859   0.1%  99.6%     2859   0.1% rocksdb::Cleanable::RegisterCleanup
    2844   0.1%  99.7%     2844   0.1% std::map::operator[] (inline)
</pre>

The "MergeValuesNewToOld (inline)" objects are the #define wrappers to SyncPoint::Process.  We discovered this in a 5.18 rocksdb release.  There TCMALLOC was more specific that std::basic_string was being constructed.  I believe that was before SyncPoint::Process was declared inline in subsequent releases.

The TCMALLOC top object list after this PR:

<pre>
Total: 104911 objects
   45090  43.0%  43.0%    45090  43.0% rocksdb::Cleanable::RegisterCleanup
   29995  28.6%  71.6%    29995  28.6% rocksdb::LRUCacheShard::Insert
   15229  14.5%  86.1%    15229  14.5% rocksdb::UncompressBlockContentsForCompressionType (inline)
    4373   4.2%  90.3%     4551   4.3% JVM_FindSignal
    2881   2.7%  93.0%     2881   2.7% rocksdb::::ReadBlockFromFile (inline)
    1162   1.1%  94.1%     1176   1.1% rocksdb::BlockFetcher::ReadBlockContents (inline)
    1036   1.0%  95.1%     1036   1.0% std::__cxx11::basic_string::_M_mutate
     869   0.8%  95.9%      869   0.8% std::vector::_M_realloc_insert (inline)
     806   0.8%  96.7%      806   0.8% SnmpAgent::GetVariables (inline)
</pre>

Pull Request resolved: https://github.com/facebook/rocksdb/pull/9023

Reviewed By: pdillinger

Differential Revision: D31610907

Pulled By: mrambacher

fbshipit-source-id: 574ff51b639dd46ad253a8e664a575f06b7cc85d
上级 ab675571
......@@ -60,7 +60,7 @@ void SyncPoint::ClearTrace() {
impl_->ClearTrace();
}
void SyncPoint::Process(const std::string& point, void* cb_arg) {
void SyncPoint::Process(const Slice& point, void* cb_arg) {
impl_->Process(point, cb_arg);
}
......
......@@ -5,6 +5,7 @@
#pragma once
#include <assert.h>
#include <functional>
#include <mutex>
#include <string>
......@@ -12,6 +13,7 @@
#include <vector>
#include "rocksdb/rocksdb_namespace.h"
#include "rocksdb/slice.h"
#ifdef NDEBUG
// empty in release build
......@@ -117,7 +119,16 @@ class SyncPoint {
// triggered by TEST_SYNC_POINT, blocking execution until all predecessors
// are executed.
// And/or call registered callback function, with argument `cb_arg`
void Process(const std::string& point, void* cb_arg = nullptr);
void Process(const Slice& point, void* cb_arg = nullptr);
// template gets length of const string at compile time,
// avoiding strlen() at runtime
template <size_t kLen>
void Process(const char (&point)[kLen], void* cb_arg = nullptr) {
static_assert(kLen > 0, "Must not be empty");
assert(point[kLen - 1] == '\0');
Process(Slice(point, kLen - 1), cb_arg);
}
// TODO: it might be useful to provide a function that blocks until all
// sync points are cleared.
......
......@@ -101,19 +101,23 @@ void SyncPoint::Data::ClearAllCallBacks() {
callbacks_.clear();
}
void SyncPoint::Data::Process(const std::string& point, void* cb_arg) {
void SyncPoint::Data::Process(const Slice& point, void* cb_arg) {
if (!enabled_) {
return;
}
// Use a filter to prevent mutex lock if possible.
if (!point_filter_.MayContain(point)) {
return;
}
// Must convert to std::string for remaining work. Take
// heap hit.
std::string point_string(point.ToString());
std::unique_lock<std::mutex> lock(mutex_);
auto thread_id = std::this_thread::get_id();
auto marker_iter = markers_.find(point);
auto marker_iter = markers_.find(point_string);
if (marker_iter != markers_.end()) {
for (auto& marked_point : marker_iter->second) {
marked_thread_id_.emplace(marked_point, thread_id);
......@@ -121,18 +125,18 @@ void SyncPoint::Data::Process(const std::string& point, void* cb_arg) {
}
}
if (DisabledByMarker(point, thread_id)) {
if (DisabledByMarker(point_string, thread_id)) {
return;
}
while (!PredecessorsAllCleared(point)) {
while (!PredecessorsAllCleared(point_string)) {
cv_.wait(lock);
if (DisabledByMarker(point, thread_id)) {
if (DisabledByMarker(point_string, thread_id)) {
return;
}
}
auto callback_pair = callbacks_.find(point);
auto callback_pair = callbacks_.find(point_string);
if (callback_pair != callbacks_.end()) {
num_callbacks_running_++;
mutex_.unlock();
......@@ -140,7 +144,7 @@ void SyncPoint::Data::Process(const std::string& point, void* cb_arg) {
mutex_.lock();
num_callbacks_running_--;
}
cleared_points_.insert(point);
cleared_points_.insert(point_string);
cv_.notify_all();
}
} // namespace ROCKSDB_NAMESPACE
......
......@@ -95,7 +95,7 @@ struct SyncPoint::Data {
return marked_point_iter != marked_thread_id_.end() &&
thread_id != marked_point_iter->second;
}
void Process(const std::string& point, void* cb_arg);
void Process(const Slice& point, void* cb_arg);
};
} // namespace ROCKSDB_NAMESPACE
#endif // NDEBUG
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册