多处理器系统下的伪共享（false sharing）问题 2 - 程序演示

原文：http://blog.youkuaiyun.com/sy8111/archive/2006/09/05/1178107.aspx

 

下面用一个简单的小程序来验证False Sharing对性能的影响。

 

这个小程序同时运行两个线程，对同一块内存区域的临近位置进行int大小的多次读写。线程1读写第1个int值，线程2读写第2个int值。按照预想，如果两个int相隔太近，落入同一个cache line，两个线程就会因为"伪共享"而花费额外的时间进行缓存一致性通信和维护，造成性能下降。

 

TestFalseSharing.cpp 是主程序，Thread.h/cpp 是对Win32 Thread API的简单封装，Timer.h/cpp是对CRT time函数的简单封装。

 

// TestFalseSharing.cpp #include "stdafx.h" #include "Thread.h" #include "Timer.h" #define COUNT 3000000000 #define DISTANCE 4 DWORD WINAPI t1_p(void* a) { int* arr = (int*)a; for (int i=0; i<COUNT; i++) arr[0] = 30; return 0; } DWORD WINAPI t2_p(void* a) { int* arr = (int*)a; for (int i=0; i<COUNT; i++) arr[DISTANCE] = 30; return 0; } int arr[64]; int main(int argc, char* argv[]) { yshi::Timer timer(std::cout); // Start threads yshi::WinThread t1(t1_p, arr); yshi::WinThread t2(t2_p, arr); // Wait for all threads to terminate t1.Wait(); t2.Wait(); timer.Stop(); return 0; }  

 

// Thread.h namespace yshi { typedef LPTHREAD_START_ROUTINE WinThreadProc; // int func_name(void*) // This class is exported from the Win32.dll class WIN32_API WinThread { public: WinThread(WinThreadProc aProc, void* aParam, bool aSuspend = false); void Resume(); void Wait(); virtual ~WinThread(); private: HANDLE _handle; private: WinThread(const WinThread&); WinThread& operator=(const WinThread&); }; }

 

// Thread.cpp #include "stdafx.h" #include "Thread.h" namespace yshi { WinThread::WinThread(WinThreadProc aProc, void* aParam, bool aSuspend) { HANDLE h = CreateThread(NULL, 0, aProc, aParam, aSuspend ? 0x04 : 0x00, NULL); if (h == NULL) { throw GetLastError(); } _handle = h; } void WinThread::Resume() { DWORD ret = ResumeThread(_handle); if (ret == -1) { throw GetLastError(); } } void WinThread::Wait() { DWORD ret = WaitForSingleObject(_handle, INFINITE); if (ret == WAIT_FAILED) throw GetLastError(); } WinThread::~WinThread() { if (_handle != 0) CloseHandle(_handle); } }  

 

// Timer.h namespace yshi { class DEBUGUTIL_API Timer { public: Timer(std::ostream& aOutput); ~Timer(); void Start(); void Stop(); private: void Print(std::ostream& aOutput); private: std::ostream& mOutput; unsigned long long mTimeDiff; private: Timer& operator=(const Timer&); }; }

 

// Timer.cpp #include "Timer.h" #include <ctime> namespace yshi { Timer::Timer(std::ostream& aOutput) : mOutput(aOutput) { Start(); } Timer::~Timer() { } void Timer::Start() { mTimeDiff = time(0); } void Timer::Stop() { mTimeDiff = time(0) - mTimeDiff; Print(mOutput); } void Timer::Print(std::ostream& aOutput) { aOutput << "Time spent: " << mTimeDiff << " seconds/n"; } }

 

运行结果

 

数据距离（DISTANCE）    花费时间（秒）

4                                 21

8                                 21

16                               10

32                               10

 

可以看到，当距离达到16个int，也就是64字节的时候，"伪共享"消失，而64字节也就是我的机器的cache line 大小。