Question

我写了一个小程序来检查线程的性能，我从我获得的结果中发现了几个问题

（我的笔记本电脑的CPU是i5 3220M）

1）每次运行程序时，所需的时间为2个线程。是因为我使用的omp计时器还是我在程序中有一些逻辑错误？

2）如果我使用cpu周期来测量性能会更好吗？

3）随着线程数量的增加，时间继续减少。我知道我的程序很简单所以可能不需要上下文切换但是额外的性能来自哪里？因为cpu调整自己的turbo freq？（根据英特尔网站，正常2.6MHz，turbo 3.3MHz）

谢谢！

输出添加1次，每次1000万次

Average Time Elapsed for 1 threads = 3.11565(Check = 5000000000)
Average Time Elapsed for 2 threads = 4.54309(Check = 5000000000)
Average Time Elapsed for 4 threads = 2.19321(Check = 5000000000)
Average Time Elapsed for 8 threads = 2.48927(Check = 5000000000)
Average Time Elapsed for 16 threads = 1.84427(Check = 5000000000)
Average Time Elapsed for 32 threads = 1.30958(Check = 5000000000)
Average Time Elapsed for 64 threads = 1.08472(Check = 5000000000)
Average Time Elapsed for 128 threads = 0.996898(Check = 5000000000)
Average Time Elapsed for 256 threads = 1.01366(Check = 5000000000)
Average Time Elapsed for 512 threads = 0.951436(Check = 5000000000)
Average Time Elapsed for 1024 threads = 0.973331(Check = 4999997440)

程序

#include <iostream>
#include <thread>
#include <algorithm>    // for_each
#include <vector>
#include <omp.h>        // omp_get_wtime

class Adder{
public:
    long sum;
    Adder(){};
    void operator()(long endVal_i){
        sum = 0;
        for (long i = 1; i<= endVal_i; i++)
            sum++;
    };
};
int main()
{
    long totalCount = 1000000000;
    int maxThread = 1025;
    int numSample = 5;

    std::vector<std::thread> threads;
    Adder adderArray[maxThread];

    std::cout << "Adding 1 for " << totalCount/1000000 << " million times\n\n";

    for (int numThread = 1; numThread <=maxThread; numThread=numThread*2){

        double avgTime=0;
        long check = 0;

        for (int i = 1; i<=numSample; i++){
            double startTime = omp_get_wtime();

            long loop = totalCount/numThread;
            for (int i = 0; i<numThread;i++)
                threads.push_back(std::thread(std::ref(adderArray[i]), loop));

            std::for_each(threads.begin(), threads.end(),std::mem_fn(&std::thread::join));

            double endTime = omp_get_wtime();

            for (int i = 0; i<numThread;i++)
                check += adderArray[i].sum;

            threads.erase(threads.begin(), threads.end());

            avgTime += endTime - startTime;

        }

        std::cout << "Average Time Elapsed for "  << numThread<< " threads = " << avgTime/numSample << "(Check = "<<check<<")\n";
    }

}

多线程的时序性能

0 个答案: