1 // Mini-benchmark for creating a lot of threads.
2 //
3 // Some facts:
4 // a) clang -O1 takes <15ms to start N=500 threads,
5 // consuming ~4MB more RAM than N=1.
6 // b) clang -O1 -ftsan takes ~26s to start N=500 threads,
7 // eats 5GB more RAM than N=1 (which is somewhat expected but still a lot)
8 // but then it consumes ~4GB of extra memory when the threads shut down!
9 // (definitely not in the barrier_wait interceptor)
10 // Also, it takes 26s to run with N=500 vs just 1.1s to run with N=1.
11 #include <assert.h>
12 #include <pthread.h>
13 #include <stdio.h>
14 #include <stdlib.h>
15 #include <unistd.h>
16
17 pthread_barrier_t all_threads_ready;
18
Thread(void * unused)19 void* Thread(void *unused) {
20 pthread_barrier_wait(&all_threads_ready);
21 return 0;
22 }
23
main(int argc,char ** argv)24 int main(int argc, char **argv) {
25 int n_threads;
26 if (argc == 1) {
27 n_threads = 100;
28 } else if (argc == 2) {
29 n_threads = atoi(argv[1]);
30 } else {
31 printf("Usage: %s n_threads\n", argv[0]);
32 return 1;
33 }
34 printf("%s: n_threads=%d\n", __FILE__, n_threads);
35
36 pthread_barrier_init(&all_threads_ready, NULL, n_threads + 1);
37
38 pthread_t *t = new pthread_t[n_threads];
39 for (int i = 0; i < n_threads; i++) {
40 int status = pthread_create(&t[i], 0, Thread, (void*)i);
41 assert(status == 0);
42 }
43 // sleep(5); // FIXME: simplify measuring the memory usage.
44 pthread_barrier_wait(&all_threads_ready);
45 for (int i = 0; i < n_threads; i++) {
46 pthread_join(t[i], 0);
47 }
48 // sleep(5); // FIXME: simplify measuring the memory usage.
49 delete [] t;
50
51 return 0;
52 }
53