今天通过酷壳一篇推荐了解了一下 Cache 的“伪共享”(False Sharing). 写了小程序做了个简单的测试：
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 | //! false_sharing.cpp #include <iostream> #include <pthread.h> #include <sys/time.h> using namespace std; template <size_t NPAD, size_t NTH = 4, size_t NLOOP = (1<<30)> class FalseSharing { public: void run() { struct timeval b, e; gettimeofday(&b, NULL); pthread_t threads[NTH]; for (size_t i = 0; i < NTH; ++i) { pthread_create(&threads[i], NULL, hook, reinterpret_cast<void*> (i)); } for (size_t i = 0; i < NTH; ++i) { pthread_join(threads[i], NULL); } gettimeofday(&e, NULL); cout<<"padding bytes: "<<NPAD<<endl;; cout<<"thread count: "<<NTH<<endl; cout<<"loop count: "<<NLOOP<<endl; cout<<"elapsed(ms): "<< ((e.tv_sec-b.tv_sec)*1000000 + (e.tv_usec-b.tv_usec)) / 1000 <<endl; cout<<endl; } private: static void* hook(void *args) { size_t ith = reinterpret_cast<size_t> (args); for (size_t i = 0; i < NLOOP; ++i) { ++s[ith].n; } return NULL; } private: struct S { size_t n; char padding[NPAD]; }; static S s[NTH]; }; template <size_t NPAD, size_t NTH, size_t NLOOP> typename FalseSharing<NPAD, NTH, NLOOP>::S FalseSharing<NPAD, NTH, NLOOP>::s[NTH]; int main() { FalseSharing<0> test1; test1.run(); FalseSharing<56> test2; test2.run(); return 0; }; |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 | $ cat /proc/cpuinfo | egrep 'model name|cache_alignment' model name : Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz cache_alignment : 64 model name : Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz cache_alignment : 64 model name : Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz cache_alignment : 64 model name : Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz cache_alignment : 64 $ g++ false_sharing.cpp -pthread $ time ./a.out padding bytes : 0 thread count : 4 loop count : 1073741824 elapsed(ms) : 12638 padding bytes : 56 thread count : 4 loop count : 1073741824 elapsed(ms) : 3888 real 0m16.530s user 1m3.688s sys 0m0.044s |
PS. 上面结果是在一台2核4线程的i7-3520笔记本上面测得的，在一台4核16线程的E5520(2.27GHz)机器上，结果更加触目惊心：
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 | padding bytes: 0 thread count: 4 loop count: 1073741824 elapsed(ms): 13646 padding bytes: 56 thread count: 4 loop count: 1073741824 elapsed(ms): 3548 padding bytes: 0 thread count: 16 loop count: 1073741824 elapsed(ms): 43848 padding bytes: 56 thread count: 16 loop count: 1073741824 elapsed(ms): 5051 |
UPDATE
有意思的是，根据Felix同学的测试性能差距并没有这么大，详情请移步到这里。