[英]c++ std library parallel execution with for_each slower than sequential loop
[英]Parallel for_each more than two times slower than std::for_each
我正在阅读Anthony Williams 撰写的C ++ Concurrency in Action 。 在有关设计并发代码的章节中,有并行版本的std :: for_each algorihtm。 这是本书中经过稍微修改的代码:
join_thread.hpp
#pragma once
#include <vector>
#include <thread>
class join_threads
{
public:
explicit join_threads(std::vector<std::thread>& threads)
: threads_(threads) {}
~join_threads()
{
for (size_t i = 0; i < threads_.size(); ++i)
{
if(threads_[i].joinable())
{
threads_[i].join();
}
}
}
private:
std::vector<std::thread>& threads_;
};
parallel_for_each.hpp
#pragma once
#include <future>
#include <algorithm>
#include "join_threads.hpp"
template<typename Iterator, typename Func>
void parallel_for_each(Iterator first, Iterator last, Func func)
{
const auto length = std::distance(first, last);
if (0 == length) return;
const auto min_per_thread = 25u;
const unsigned max_threads = (length + min_per_thread - 1) / min_per_thread;
const auto hardware_threads = std::thread::hardware_concurrency();
const auto num_threads= std::min(hardware_threads != 0 ?
hardware_threads : 2u, max_threads);
const auto block_size = length / num_threads;
std::vector<std::future<void>> futures(num_threads - 1);
std::vector<std::thread> threads(num_threads-1);
join_threads joiner(threads);
auto block_start = first;
for (unsigned i = 0; i < num_threads - 1; ++i)
{
auto block_end = block_start;
std::advance(block_end, block_size);
std::packaged_task<void (void)> task([block_start, block_end, func]()
{
std::for_each(block_start, block_end, func);
});
futures[i] = task.get_future();
threads[i] = std::thread(std::move(task));
block_start = block_end;
}
std::for_each(block_start, last, func);
for (size_t i = 0; i < num_threads - 1; ++i)
{
futures[i].get();
}
}
我使用以下程序使用顺序版本的std :: for_each对它进行了基准测试:
main.cpp
#include <iostream>
#include <random>
#include <chrono>
#include "parallel_for_each.hpp"
using namespace std;
constexpr size_t ARRAY_SIZE = 500'000'000;
typedef std::vector<uint64_t> Array;
template <class FE, class F>
void test_for_each(const Array& a, FE fe, F f, atomic<uint64_t>& result)
{
auto time_begin = chrono::high_resolution_clock::now();
result = 0;
fe(a.begin(), a.end(), f);
auto time_end = chrono::high_resolution_clock::now();
cout << "Result = " << result << endl;
cout << "Time: " << chrono::duration_cast<chrono::milliseconds>(
time_end - time_begin).count() << endl;
}
int main()
{
random_device device;
default_random_engine engine(device());
uniform_int_distribution<uint8_t> distribution(0, 255);
Array a;
a.reserve(ARRAY_SIZE);
cout << "Generating array ... " << endl;
for (size_t i = 0; i < ARRAY_SIZE; ++i)
a.push_back(distribution(engine));
atomic<uint64_t> result;
auto acc = [&result](uint64_t value) { result += value; };
cout << "parallel_for_each ..." << endl;
test_for_each(a, parallel_for_each<Array::const_iterator, decltype(acc)>, acc, result);
cout << "for_each ..." << endl;
test_for_each(a, for_each<Array::const_iterator, decltype(acc)>, acc, result);
return 0;
}
我的机器上算法的并行版本比顺序算法慢两倍以上:
parallel_for_each ...
Result = 63750301073
Time: 5448
for_each ...
Result = 63750301073
Time: 2496
我在运行Intel(R)Core(TM)i3-6100 CPU @ 3.70GHz的 Ubuntu Linux上使用GCC 6.2编译器。
如何解释这种行为? 这是因为线程和缓存乒乓之间共享atomic<uint64_t>
变量吗?
我分别用perf进行了分析 。 对于并行版本,统计信息如下:
1137982167 cache-references
247652893 cache-misses # 21,762 % of all cache refs
60868183996 cycles
27409239189 instructions # 0,45 insns per cycle
3287117194 branches
80895 faults
4 migrations
对于顺序一:
402791485 cache-references
246561299 cache-misses # 61,213 % of all cache refs
40284812779 cycles
26515783790 instructions # 0,66 insns per cycle
3188784664 branches
48179 faults
3 migrations
显然,并行版本会生成更多的缓存引用,周期和错误,但是为什么呢?
您正在共享同一个result
变量 :所有线程都在atomic<uint64_t> result
上累积, atomic<uint64_t> result
破坏了缓存!
每当线程写入result
,其他内核中的所有缓存都会失效:这会导致缓存行争用 。
更多信息:
要写入存储位置,内核还必须另外拥有包含该位置的缓存行的专有所有权。 虽然一个内核可以独占使用,但是其他所有试图写入同一内存位置的内核都必须等待并轮流使用-也就是说,它们必须串行运行。 从概念上讲,每条高速缓存行好像都受到硬件互斥锁的保护,在这种情况下,一次只有一个内核可以将硬件锁持有在该高速缓存行上。
关于“错误共享”的这篇文章涵盖了类似的问题,它更深入地解释了缓存中发生的情况。
我对您的程序进行了一些修改,并获得了以下结果(在具有i7-4770K [8个线程+超线程]的计算机上) :
Generating array ...
parallel_for_each ...
Result = 63748111806
Time: 195
for_each ...
Result = 63748111806
Time: 2727
并行版本比串行版本快大约92% 。
std::future
和std::packaged_task
是重量级抽象。 在这种情况下,一个std::experimental::latch
就足够了。
每个任务都被发送到线程池。这样可以最大程度地减少线程创建的开销 。
每个任务都有自己的累加器 。 这消除了共享 。
该代码可在我的GitHub上找到 。 它使用了一些个人依赖项,但是无论如何您都应该了解所做的更改。
以下是最重要的更改:
// A latch is being used instead of a vector of futures.
ecst::latch l(num_threads - 1);
l.execute_and_wait_until_zero([&]
{
auto block_start = first;
for (unsigned i = 0; i < num_threads - 1; ++i)
{
auto block_end = block_start;
std::advance(block_end, block_size);
// `p` is a thread pool.
// Every task posted in the thread pool has its own `tempacc` accumulator.
p.post([&, block_start, block_end, tempacc = 0ull]() mutable
{
// The task accumulator is filled up...
std::for_each(block_start, block_end, [&tempacc](auto x){ tempacc += x; });
// ...and then the atomic variable is incremented ONCE.
func(tempacc);
l.decrement_and_notify_all();
});
block_start = block_end;
}
// Same idea here: accumulate to local non-atomic counter, then
// add the partial result to the atomic counter ONCE.
auto tempacc2 = 0ull;
std::for_each(block_start, last, [&tempacc2](auto x){ tempacc2 += x; });
func(tempacc2);
});
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.