int main(int argc, char **argv)
{
int i;
pthread_t *id;
double start, time = 0;
process_args(argc,argv);
// bind_to_core(0);
pthread_barrier_init(&tbarrier, NULL, nthreads + 2);
id = calloc(nthreads + 1, sizeof(*id));
list = tslist_create(nthreads, nadds * nbatch);
/* setup and create threads */
pthread_create(&id[nthreads], NULL, consumer, (void *)NULL);
for (i=0; i<nthreads; i++) {
pthread_create(&id[i], NULL, producer, (void *)NULL);
}
pthread_barrier_wait(&tbarrier);
start = GET_TS();
for (i=0; i<nthreads + 1; i++) {
pthread_join(id[i], NULL);
}
time = GET_TS() - start;
int count = nthreads * nadds * nbatch;
printf("Number of elements is %d\n", count);
printf("Time: %lf\n", time * 1E6);
printf("Performance: %lf Melem/s\n", count / time / 1E6);
tslist_append_done(list, nthreads);
}
void *f(void* thr_data)
{
int my_idx = *(int*)thr_data;
cpu_set_t set;
CPU_ZERO(&set);
CPU_SET(my_idx, &set);
if( pthread_setaffinity_np(pthread_self(), sizeof(set), &set) ){
abort();
}
while( !start );
timings[my_idx][0] = GET_TS();
for(int n = 0; n < niter; ++n) {
_Bool flag = false;
while( !flag ){
int tmp = acnt;
flag = atomic_compare_exchange_weak(&acnt, &tmp, tmp + 1);
}
++cnt; // undefined behavior, in practice some updates missed
}
timings[my_idx][1] = GET_TS();
return 0;
}
static int run_test(int server)
{
int i;
if( server ){
connect_server();
} else {
connect_client();
}
poll_prepare();
for(i=1; i <= 4194304; i *= 2){
int rep, j;
if( i < 4096 ){
rep = 100;
} else {
rep = 10;
}
/* warmup */
for(j=0; j < rep/10; j++){
progress(server, i);
}
double start = GET_TS();
for(j=0; j < rep; j++){
progress(server, i);
}
double time = GET_TS() - start;
printf("%d\t%lf\t%lf\n",i, time / rep / 2 * 1000000,
(double)progress_count / progress_calls);
progress_count = 0;
progress_calls = 0;
}
}
int main(int argc, char **argv)
{
if( argc < 3 ){
printf("Want <nthr> and <niter>\n");
return 0;
}
nthr = atoi(argv[1]);
niter = atoi(argv[2]);
pthread_t thr[nthr];
for(int n = 0; n < nthr; ++n){
indexes[n] = n;
pthread_create(&thr[n], NULL, f, &indexes[n]);
}
start = 1;
for(int n = 0; n < nthr; ++n)
pthread_join(thr[n], NULL);
double min_start = GET_TS(), max_start = 0, avg_start = 0;
double min_work = GET_TS(), max_work = 0, avg_work = 0;
int min_w_idx = -1, max_w_idx = -1;
for(int n=0; n < nthr; n++){
if( min_start > timings[n][0] ){
min_start = timings[n][0];
}
avg_start += timings[n][0];
if( max_start < timings[n][0] ){
max_start = timings[n][0];
}
double interval = timings[n][1] - timings[n][0];
if( interval < min_work ){
min_work = interval;
min_w_idx = n;
}
if( interval > max_work ){
max_w_idx = n;
max_work = interval;
}
avg_work += interval;
}
printf("The atomic counter is %u\n", acnt);
printf("The non-atomic counter is %u\n", cnt);
printf("Start: %lf/%lf/%lf\n", avg_start / nthr, min_start, max_start);
printf("Work : %lf / %lf / %lf %d/%d\n", avg_work / nthr, min_work, max_work, min_w_idx, max_w_idx);
return 0;
}
bool pmixp_server_pp_check_fini(int size)
{
if ( (pmixp_server_pp_count() + 1) >=
(_pmixp_pp_warmup + _pmixp_pp_iters)){
slurm_mutex_lock(&_pmixp_pp_lock);
PMIXP_ERROR("latency: %d - %.9lf", size,
(GET_TS() - _pmixp_pp_start) / _pmixp_pp_iters );
slurm_mutex_unlock(&_pmixp_pp_lock);
return true;
}
return false;
}
void pmixp_server_pp_start(void)
{
_pmixp_pp_start = GET_TS();
}
