锁的代价

秃子说并发程序的性能瓶颈常常在于锁的滥用。我对此虽深信不疑，却一直没有量化的感性认识。刚才写了个浮点运算小程序一跑，结果显示对一次浮点加法来说，有锁无锁性能相差8到10倍。注意，测试程序是单线程的，也就是说这还只是没有实际竞争时的差别。
测试代码如下：
#include <pthread.h> #include <time.h> #include <stdio.h> #include <stdlib.h> #include <math.h> class RealMutex { public: RealMutex() { pthread_mutex_init(&mutex, 0); } ~RealMutex() { pthread_mutex_destroy(&mutex); } public: void lock() { pthread_mutex_lock(&mutex); } void unlock() { pthread_mutex_unlock(&mutex); } private: pthread_mutex_t mutex; }; class FakeMutex { public: void lock() {} void unlock() {} }; template<class MutexT> void test() { const int LOOP_COUNT = 10000 * 10000; MutexT mutex; double d = 0.0; clock_t start = clock(); for(int i = 0; i < LOOP_COUNT; ++i) { mutex.lock(); d += 1.0; mutex.unlock(); } clock_t end = clock(); if(fabs(d - LOOP_COUNT) > 1e-8) exit(-1); clock_t ms = (end - start) * 1000 / CLOCKS_PER_SEC; printf("Time consumed: %ld (ms)./n", ms); } int main() { test<RealMutex>(); test<FakeMutex>(); return 0; }

在我的机器上输出如下：
g++默认：
Time consumed: 4230 (ms).
Time consumed: 520 (ms).
g++ -0:
Time consumed: 3360 (ms).
Time consumed: 340 (ms).

机器环境：
Linux 2.6.32-24-server #39-Ubuntu SMP x86_64 GNU/Linux
Intel(R) Core(TM)2 Quad CPU    Q9400  @ 2.66GHz
gcc (Ubuntu 4.4.3-4ubuntu5) 4.4.3