int MTest_thread_barrier_init(void)
{
bcount = -1; /* must reset to force barrier re-creation */
return MTest_thread_lock_create(&barrierLock);
}
int MTest_thread_barrier_init(void)
{
return MTest_thread_lock_create(&barrierLock);
}
void loops(void)
{
int i, nt;
double latency, mrate, avg_latency, agg_mrate;
int err;
err = MTest_thread_lock_create(&num_threads_lock);
if (err) ABORT_MSG("unable to create lock, aborting\n");
for (nt = 1; nt <= MAX_THREADS; nt++) {
err = MTest_thread_lock(&num_threads_lock);
if (err) ABORT_MSG("unable to acquire lock, aborting\n");
num_threads = 1;
MPI_Barrier(MPI_COMM_WORLD);
MTest_thread_barrier_init();
for (i = 1; i < nt; i++) {
err = MTest_Start_thread(run_test, (void *)(long)i);
if (err) {
/* attempt to continue with fewer threads, we may be on a
* thread-constrained platform like BG/P in DUAL mode */
break;
}
++num_threads;
}
err = MTest_thread_unlock(&num_threads_lock);
if (err) ABORT_MSG("unable to release lock, aborting\n");
if (nt > 1 && num_threads <= 1) {
ABORT_MSG("unable to create any additional threads, aborting\n");
}
run_test((void *) 0); /* we are thread 0 */
err = MTest_Join_threads();
if (err) {
printf("error joining threads, err=%d", err);
MPI_Abort(MPI_COMM_WORLD, 1);
}
MTest_thread_barrier_free();
latency = 0;
for (i = 0; i < num_threads; i++)
latency += tp[i].latency;
latency /= num_threads; /* Average latency */
mrate = num_threads / latency; /* Message rate */
/* Global latency and message rate */
MPI_Reduce(&latency, &avg_latency, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
avg_latency /= size;
MPI_Reduce(&mrate, &agg_mrate, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
if (!rank && verbose) {
printf("Threads: %d; Latency: %.3f; Mrate: %.3f\n",
num_threads, latency, mrate);
}
}
err = MTest_thread_lock_free(&num_threads_lock);
if (err) ABORT_MSG("unable to free lock, aborting\n");
}