一个容易被大家忽略的多线程性能优化方案：Cache行级别优化

#include
#include
using namespace std;

struct Count
{
    long count1;
    char space[64];  // 此处增加64字节的“无用”空间，目的就是使得count1和count2位于不同的Cache行，从而避免Cache行同步引起的“乒乓效应”，可以尝试去掉这64字节的“无用”空间，将会得到令人诧异的运行结果，要探明深层原因，可以搜索一下“乒乓效应 CPU cache”

    long count2;
};
DWORD WINAPI Calculate1(void *p)
{
    double x = 0,y = 0;
    long* c1 = (long*)p;
    for(; x <= 0.5;x += 0.0001)
        for(y = 0;y <= 1;y += 0.0001)
        {
            if (x*x + y*y <= 1) (*c1)++;
        }
    return 0;
}
DWORD WINAPI Calculate2(void *p)
{
    double i = 0,j = 0;//16
    long* c2 = (long*)p;
    for(i=0.5001;i <=1;i+=0.0001)
        for(j=0;j <=1;j+=0.0001)
        {
            if(i*i+j*j <=1) (*c2)++;
        }
    return 0;
}

int main()
{
    DWORD pThreadID;
    Count cnt;

    DWORD startTime = 0,endTime = 0;
    startTime= GetTickCount();
    Calculate1(&cnt.count1);
    Calculate2(&cnt.count2);
    cout << "Serial calculate cost " << GetTickCount() - startTime << "ms." << endl;

    cnt.count1 = 0;
    cnt.count2 = 0;
    startTime= GetTickCount();

    HANDLE hThread1 = CreateThread (NULL,
                                    0,
                                    Calculate1,
                                    &cnt.count1,
                                    0,
                                    &pThreadID);

    HANDLE hThread2 = CreateThread (NULL,
                                    0,
                                    Calculate2,
                                    &cnt.count2,
                                    0,
                                    &pThreadID);

    WaitForSingleObject(hThread1, INFINITE);
    WaitForSingleObject(hThread2, INFINITE);
    endTime=GetTickCount();

    cout << "Parallel calculate cost " << endTime-startTime << "ms." << endl;
}

在本人的双核机子上，此程序运算结果符合双线程比单线程性能提升一倍的预期，但如果将“char space[64]”去掉，两者性能相差无几。

来自 “ ITPUB博客 ” ，链接：http://blog.itpub.net/7364811/viewspace-663783/，如需转载，请注明出处，否则将追究法律责任。

转载于:http://blog.itpub.net/7364811/viewspace-663783/