在这个程序中，我尝试测量将项目推送到并发队列与另一个线程弹出项目之间的平均时间：
#include <chrono>
#include <iomanip>
#include <iostream>
#include <thread>
#include "tbb/concurrent_queue.h"
namespace {
void test_unbounded_queue(uint32_t iteration_count)
{
typedef std::chrono::high_resolution_clock Clock;
using std::chrono::duration_cast;
using std::chrono::nanoseconds;
tbb::concurrent_queue<uint32_t> unbounded_queue;
uint32_t verify = 0;
Clock::time_point start_time, end_time;
std::thread consumer_thread([&]
{
start_time = Clock::now();
// Active polling (100% CPU usage)
while (true)
{
uint32_t n;
if (unbounded_queue.try_pop(n))
{
if (n == uint32_t(-1))
{
end_time = Clock::now();
return;
}
verify += n;
}
}
});
for (unsigned i = 0; i != iteration_count; ++i)
{
unbounded_queue.push(1);
}
// stop the thread
unbounded_queue.push(uint32_t(-1));
consumer_thread.join();
// verify the iteration count
if (verify != iteration_count)
{
std::cout << "verify: " << verify << " != " << iteration_count << std::endl;
return;
}
// print the results
auto total_time_ns = duration_cast<nanoseconds>(end_time - start_time).count();
auto cost_per_pop = total_time_ns / iteration_count;
std::cout
<< "iterations: "
<< std::setw(6) << iteration_count
<< ", total time: "
<< std::setw(9) << total_time_ns << " ns"
<< ", average time: "
<< std::setw(5) << cost_per_pop << " ns"
<< std::endl;
}
void test_bounded_queue(uint32_t iteration_count)
{
typedef std::chrono::high_resolution_clock Clock;
using std::chrono::duration_cast;
using std::chrono::nanoseconds;
tbb::concurrent_bounded_queue<uint32_t> bounded_queue;
uint32_t verify = 0;
Clock::time_point start_time, end_time;
std::thread consumer_thread([&]
{
start_time = Clock::now();
while (true)
{
uint32_t n;
bounded_queue.pop(n); // waits, no CPU usage
if (n == uint32_t(-1))
{
end_time = Clock::now();
return;
}
verify += n;
}
});
for (unsigned i = 0; i != iteration_count; ++i)
{
bounded_queue.push(1);
}
// stop the thread
bounded_queue.push(uint32_t(-1));
consumer_thread.join();
// verify the iteration count
if (verify != iteration_count)
{
std::cout << "verify: " << verify << " != " << iteration_count << std::endl;
return;
}
// print the results
auto total_time_ns = duration_cast<nanoseconds>(end_time - start_time).count();
auto cost_per_pop = total_time_ns / iteration_count;
std::cout
<< "iterations: "
<< std::setw(6) << iteration_count
<< ", total time: "
<< std::setw(9) << total_time_ns << " ns"
<< ", average time: "
<< std::setw(5) << cost_per_pop << " ns"
<< std::endl;
}
const uint32_t min_iterations = 1u;
const uint32_t max_iterations = 128u * 1024u;
void test_bounded()
{
std::cout << "Test bounded queue" << std::endl;
for (uint32_t i = min_iterations; i <= max_iterations; i *= 2)
{
test_bounded_queue(i);
}
std::cout << std::endl;
}
void test_unbounded()
{
std::cout << "Test unbounded queue" << std::endl;
for (uint32_t i = min_iterations; i <= max_iterations; i *= 2)
{
test_unbounded_queue(i);
}
std::cout << std::endl;
}
} // namespace
int main()
{
test_bounded();
test_unbounded();
}
输出如下所示：
+ g++ -std=c++11 -O2 -Wall -pedantic -pthread main.cpp -ltbb
+ ./a.out
Test bounded queue
iterations: 1, total time: 4861 ns, average time: 4861 ns
iterations: 2, total time: 2901 ns, average time: 1450 ns
iterations: 4, total time: 3111 ns, average time: 777 ns
iterations: 8, total time: 3002 ns, average time: 375 ns
iterations: 16, total time: 3475 ns, average time: 217 ns
iterations: 32, total time: 4514 ns, average time: 141 ns
iterations: 64, total time: 6329 ns, average time: 98 ns
iterations: 128, total time: 12952 ns, average time: 101 ns
iterations: 256, total time: 53654 ns, average time: 209 ns
iterations: 512, total time: 178558 ns, average time: 348 ns
iterations: 1024, total time: 395186 ns, average time: 385 ns
iterations: 2048, total time: 886728 ns, average time: 432 ns
iterations: 4096, total time: 1865523 ns, average time: 455 ns
iterations: 8192, total time: 3535521 ns, average time: 431 ns
iterations: 16384, total time: 7044609 ns, average time: 429 ns
iterations: 32768, total time: 15743366 ns, average time: 480 ns
iterations: 65536, total time: 32278651 ns, average time: 492 ns
iterations: 131072, total time: 62466063 ns, average time: 476 ns
Test unbounded queue
iterations: 1, total time: 2952 ns, average time: 2952 ns
iterations: 2, total time: 2378 ns, average time: 1189 ns
iterations: 4, total time: 2463 ns, average time: 615 ns
iterations: 8, total time: 2874 ns, average time: 359 ns
iterations: 16, total time: 3094 ns, average time: 193 ns
iterations: 32, total time: 3542 ns, average time: 110 ns
iterations: 64, total time: 4138 ns, average time: 64 ns
iterations: 128, total time: 6380 ns, average time: 49 ns
iterations: 256, total time: 14647 ns, average time: 57 ns
iterations: 512, total time: 103661 ns, average time: 202 ns
iterations: 1024, total time: 275176 ns, average time: 268 ns
iterations: 2048, total time: 564442 ns, average time: 275 ns
iterations: 4096, total time: 1182691 ns, average time: 288 ns
iterations: 8192, total time: 2278423 ns, average time: 278 ns
iterations: 16384, total time: 4767694 ns, average time: 290 ns
iterations: 32768, total time: 9382879 ns, average time: 286 ns
iterations: 65536, total time: 18611420 ns, average time: 283 ns
iterations: 131072, total time: 34437811 ns, average time: 262 ns
对于大量迭代，有界队列的平均时间平均为 400-500 ns，无界队列的平均时间为 200-300 ns。
对于低迭代次数，我预计“热身”开销会增加平均时间。然而，有时它只有~50 ns。
怎么会这样？