About four years ago, I wrote the condition variables in Win32 API was much faster than boost threading library implementation. Since then the boost threading library has been rewritten and C++11 has introduced the threading support in the standard library. Let’s revisit the benchmark again.
The test program is bounded FIFO implemented with two condition variables. It passes 10,000,000 integers from the producer to the consumer. The test is conducted 50 times with the following environment.
– Intel(R) Core (TM) i7 950@3GHz.
– Windows 7 Professional
– Visual Studio 2012 Express (Update 1)
The average speed is shown here. Shorter is faster. Notice that it’s the result of 10 million times FIFO access. Although std::condition_variable is slower than others, the difference is pretty much nothing.
However, the distributions of elapsed time are still interesting. Among 50 times trial, Win32 and boost condition variables have pretty stable performance.
For unknown reason, however, std::condition_variable of Visual Studio 2012 Update 1 implementation sometimes takes very long time.
Here is the test program.
#include <boost/thread.hpp>
#include <boost/timer.hpp>
#include <deque>
#include <thread>
#include <mutex>
#include <condition_variable>
#include <windows.h>
using namespace std;
class fifo
{
public:
virtual ~fifo(){}
virtual void push(int v) = 0;
virtual int pull() = 0;
};
class boost_fifo : public fifo
{
public:
boost_fifo(std::size_t s) : max_fifo_size(s){}
void push(int v)
{
boost::mutex::scoped_lock lk(mtx);
while(buffer.size() > max_fifo_size)
{
cv_slot.wait(lk);
}
buffer.push_back(v);
cv_data.notify_one();
}
int pull()
{
boost::mutex::scoped_lock lk(mtx);
while(buffer.empty())
{
cv_data.wait(lk);
}
int v = buffer.front();
buffer.pop_front();
cv_slot.notify_one();
return v;
}
private:
std::size_t max_fifo_size;
boost::mutex mtx;
boost::condition_variable cv_slot;
boost::condition_variable cv_data;
std::deque<int> buffer;
};
class win32_fifo : public fifo
{
public:
win32_fifo(std::size_t s) : max_fifo_size(s)
{
InitializeCriticalSection(&mtx);
InitializeConditionVariable(&cv_slot);
InitializeConditionVariable(&cv_data);
}
~win32_fifo()
{
DeleteCriticalSection(&mtx);
}
void push(int v)
{
EnterCriticalSection(&mtx);
while(buffer.size() > max_fifo_size)
{
SleepConditionVariableCS(&cv_slot, &mtx, INFINITE);
}
buffer.push_back(v);
WakeConditionVariable(&cv_data);
LeaveCriticalSection(&mtx);
}
int pull()
{
EnterCriticalSection(&mtx);
while(buffer.empty())
{
SleepConditionVariableCS(&cv_data, &mtx, INFINITE);
}
int v = buffer.front();
buffer.pop_front();
WakeConditionVariable(&cv_slot);
LeaveCriticalSection(&mtx);
return v;
}
private:
std::size_t max_fifo_size;
CRITICAL_SECTION mtx;
CONDITION_VARIABLE cv_slot;
CONDITION_VARIABLE cv_data;
std::deque<int> buffer;
};
class std_fifo : public fifo
{
public:
std_fifo(std::size_t s) : max_fifo_size(s){}
void push(int v)
{
std::unique_lock<std::mutex> lk(mtx);
while(buffer.size() > max_fifo_size)
{
cv_slot.wait(lk);
}
buffer.push_back(v);
cv_data.notify_one();
}
int pull()
{
std::unique_lock<std::mutex> lk(mtx);
while(buffer.empty())
{
cv_data.wait(lk);
}
int v = buffer.front();
buffer.pop_front();
cv_slot.notify_one();
return v;
}
private:
std::size_t max_fifo_size;
std::mutex mtx;
std::condition_variable cv_slot;
std::condition_variable cv_data;
std::deque<int> buffer;
};
void push_loop(fifo* fifo_buffer)
{
for(int i = 0; i < 10000000; i++)
{
fifo_buffer->push(i);
}
}
void pull_loop(fifo* fifo_buffer)
{
for(int i = 0; i < 10000000; i++)
{
fifo_buffer->pull();
}
}
int main()
{
vector<double> elapsedTime[3];
static size_t const FIFO_SIZE = 16;
for(int i = 0; i < 50; i++)
{
cerr << "BOOST FIFO: " << endl;
{
boost::timer tim;
unique_ptr<fifo> fifo_buffer(new boost_fifo(FIFO_SIZE));
boost::thread push_thread(bind(push_loop, fifo_buffer.get()));
boost::thread pull_thread(bind(pull_loop, fifo_buffer.get()));
push_thread.join();
pull_thread.join();
elapsedTime[0].push_back(tim.elapsed());
}
cerr << "WIN32 FIFO: " << endl;
{
boost::timer tim;
unique_ptr<fifo> fifo_buffer(new win32_fifo(FIFO_SIZE));
boost::thread push_thread(bind(push_loop, fifo_buffer.get()));
boost::thread pull_thread(bind(pull_loop, fifo_buffer.get()));
push_thread.join();
pull_thread.join();
elapsedTime[1].push_back(tim.elapsed());
}
cerr << "STD FIFO: " << endl;
{
boost::timer tim;
unique_ptr<fifo> fifo_buffer(new std_fifo(FIFO_SIZE));
boost::thread push_thread(bind(push_loop, fifo_buffer.get()));
boost::thread pull_thread(bind(pull_loop, fifo_buffer.get()));
push_thread.join();
pull_thread.join();
elapsedTime[2].push_back(tim.elapsed());
}
}
for(auto v : elapsedTime)
{
for(auto e : v)
{
cout << e << ", ";
}
cout << endl;
}
return 1;
}
I used the following R code generate the graph. It also contains the raw data of elapsed time.
std <- c(5.183, 5.701, 7.006, 5.839, 5.524, 16.463, 8.336, 7.041,
5.574, 5.098, 5.430, 5.034, 5.244, 6.088, 7.478, 5.557,
5.062, 5.577, 5.266, 5.991, 5.900, 6.804, 6.452, 6.981,
10.586, 7.152, 37.248, 5.469, 12.113, 6.645, 8.030, 10.559,
10.695, 6.878, 6.321, 11.659, 6.210, 17.335, 5.804, 6.146,
5.067, 5.709, 5.938, 5.718, 8.967, 7.143, 5.078, 7.472,
5.347, 15.000)
boost <- c(6.104, 6.11, 5.862, 5.07, 5.311, 5.639, 5.402, 5.532, 5.61,
5.686, 6.129, 5.481, 5.07, 5.602, 5.276, 6.046, 5.818, 6.209,
6.111, 6.165, 5.828, 6.266, 5.806, 6.087, 5.708, 6.119, 6.406,
6.096, 6.005, 6.332, 6.522, 6.91, 6.921, 5.765, 6.223, 5.444,
6.274, 6.603, 5.717, 6.471, 5.968, 6.188, 6.608, 5.333, 5.506,
5.507, 6.241, 5.981, 5.915, 6.515)
win32 <- c(4.356, 5.354, 4.641, 7.821, 4.664, 4.593, 7.174, 5.596, 4.493,
4.353, 8.678, 4.459, 5.135, 4.526, 5.043, 4.997, 5.148, 5.03,
7.421, 6.344, 7.64, 5.491, 6.101, 6.054, 6.626, 6.292, 6.29,
6.912, 5.691, 5.991, 6.263, 6.096, 6.544, 8.024, 6.405, 6.216,
6.157, 5.821, 7.759, 5.454, 6.364, 8.168, 8.471, 5.108, 6.192,
5.385, 5.427, 5.855, 6.099, 5.539)
avg <- c(mean(std), mean(boost), mean(win32))
png("cv_comp_average.png")
barplot(avg, main="Condition Variable Comparison", names.arg = c("STD", "boost", "win32"), ylab="Elapsed Time (seconds)", ylim=c(0, 10))
dev.off()
png("cv_comp_std_dist.png")
hist(std, main="STD elapsed time histgram", xlab="Elapsed Time (seconds)", breaks=seq(0, 50, 1.0), ylim=c(0, 30))
dev.off()
png("cv_comp_boost_dist.png")
hist(boost, main="boost elapsed time histgram", xlab="Elapsed Time (seconds)", breaks=seq(0, 50, 1.0), ylim=c(0, 30))
dev.off()
png("cv_comp_win32_dist.png")
hist(boost, main="win32 elapsed time histgram", xlab="Elapsed Time (seconds)", breaks=seq(0, 50, 1.0), ylim=c(0, 30))
dev.off()
. Also suppose each developer downloads the source tree 4 times a day, which is 100 times per week as 
. 
, 
, and 
, thus
, which is 
, the waste ratio is about 
, which is 1.56%.
