ucs_status_t ucx_perf_run(ucx_perf_params_t *params, ucx_perf_result_t *result)
{
ucx_perf_context_t *perf;
ucs_status_t status;
if (params->command == UCX_PERF_CMD_LAST) {
ucs_error("Test is not selected");
status = UCS_ERR_INVALID_PARAM;
goto out;
}
if ((params->api != UCX_PERF_API_UCT) && (params->api != UCX_PERF_API_UCP)) {
ucs_error("Invalid test API parameter (should be UCT or UCP)");
status = UCS_ERR_INVALID_PARAM;
goto out;
}
perf = malloc(sizeof(*perf));
if (perf == NULL) {
status = UCS_ERR_NO_MEMORY;
goto out;
}
ucx_perf_test_reset(perf, params);
status = ucx_perf_funcs[params->api].setup(perf, params);
if (status != UCS_OK) {
goto out_free;
}
if (UCS_THREAD_MODE_SINGLE == params->thread_mode) {
if (params->warmup_iter > 0) {
ucx_perf_set_warmup(perf, params);
status = ucx_perf_funcs[params->api].run(perf);
if (status != UCS_OK) {
goto out_cleanup;
}
rte_call(perf, barrier);
ucx_perf_test_reset(perf, params);
}
/* Run test */
status = ucx_perf_funcs[params->api].run(perf);
rte_call(perf, barrier);
if (status == UCS_OK) {
ucx_perf_calc_result(perf, result);
rte_call(perf, report, result, perf->params.report_arg, 1);
}
} else {
status = ucx_perf_thread_spawn(perf, result);
}
out_cleanup:
ucx_perf_funcs[params->api].cleanup(perf);
out_free:
free(perf);
out:
return status;
}
static int ucx_perf_thread_spawn(ucx_perf_params_t* params,
ucx_perf_result_t* result) {
ucx_perf_context_t perf;
ucs_status_t status;
int ti;
int nti = params->thread_count;
ucx_perf_thread_context_t* tctx =
calloc(nti, sizeof(ucx_perf_thread_context_t));
ucs_status_t* statuses =
calloc(nti, sizeof(ucs_status_t));
pthread_barrier_t tbarrier;
pthread_barrier_init(&tbarrier, NULL, nti);
ucx_perf_test_reset(&perf, params);
status = ucx_perf_funcs[params->api].setup(&perf, params);
if (UCS_OK != status) {
goto out_cleanup;
}
for (ti = 0; ti < nti; ti++) {
tctx[ti].tid = ti;
tctx[ti].ntid = nti;
tctx[ti].tbarrier = &tbarrier;
tctx[ti].statuses = statuses;
tctx[ti].params = *params;
tctx[ti].perf = perf;
/* Doctor the src and dst buffers to make them thread specific */
tctx[ti].perf.send_buffer += ti * params->message_size;
tctx[ti].perf.recv_buffer += ti * params->message_size;
pthread_create(&tctx[ti].pt, NULL,
ucx_perf_thread_run_test, (void*)&tctx[ti]);
}
for (ti = 0; ti < nti; ti++) {
pthread_join(tctx[ti].pt, NULL);
if (UCS_OK != statuses[ti]) {
ucs_error("Thread %d failed to run test: %s", tctx[ti].tid, ucs_status_string(statuses[ti]));
status = statuses[ti];
}
}
ucx_perf_funcs[params->api].cleanup(&perf);
out_cleanup:
pthread_barrier_destroy(&tbarrier);
free(statuses);
free(tctx);
return status;
}
static int ucx_perf_thread_spawn(ucx_perf_params_t* params,
ucx_perf_result_t* result) {
ucx_perf_context_t perf;
ucs_status_t status = UCS_OK;
int ti, nti;
omp_set_num_threads(params->thread_count);
nti = params->thread_count;
ucx_perf_thread_context_t* tctx =
calloc(nti, sizeof(ucx_perf_thread_context_t));
ucs_status_t* statuses =
calloc(nti, sizeof(ucs_status_t));
ucx_perf_test_reset(&perf, params);
status = ucx_perf_funcs[params->api].setup(&perf, params);
if (UCS_OK != status) {
goto out_cleanup;
}
#pragma omp parallel private(ti)
{
ti = omp_get_thread_num();
tctx[ti].tid = ti;
tctx[ti].ntid = nti;
tctx[ti].statuses = statuses;
tctx[ti].params = *params;
tctx[ti].perf = perf;
/* Doctor the src and dst buffers to make them thread specific */
tctx[ti].perf.send_buffer += ti * params->message_size;
tctx[ti].perf.recv_buffer += ti * params->message_size;
tctx[ti].perf.offset = ti * params->message_size;
ucx_perf_thread_run_test((void*)&tctx[ti]);
}
for (ti = 0; ti < nti; ti++) {
if (UCS_OK != statuses[ti]) {
ucs_error("Thread %d failed to run test: %s", tctx[ti].tid, ucs_status_string(statuses[ti]));
status = statuses[ti];
}
}
ucx_perf_funcs[params->api].cleanup(&perf);
out_cleanup:
free(statuses);
free(tctx);
return status;
}
static void* ucx_perf_thread_run_test(void* arg) {
ucx_perf_thread_context_t* tctx = (ucx_perf_thread_context_t*) arg;
ucx_perf_params_t* params = &tctx->params;
ucx_perf_result_t* result = &tctx->result;
ucx_perf_context_t* perf = &tctx->perf;
ucs_status_t* statuses = tctx->statuses;
pthread_barrier_t* tbarrier = tctx->tbarrier;
int tid = tctx->tid;
int i;
if (params->warmup_iter > 0) {
ucx_perf_set_warmup(perf, params);
statuses[tid] = ucx_perf_funcs[params->api].run(perf);
pthread_barrier_wait(tbarrier);
for (i = 0; i < tctx->ntid; i++) {
if (UCS_OK != statuses[i]) {
goto out;
}
}
if (0 == tid) {
rte_call(perf, barrier);
ucx_perf_test_reset(perf, params);
}
}
/* Run test */
pthread_barrier_wait(tbarrier);
statuses[tid] = ucx_perf_funcs[params->api].run(perf);
pthread_barrier_wait(tbarrier);
for (i = 0; i < tctx->ntid; i++) {
if (UCS_OK != statuses[i]) {
goto out;
}
}
if (0 == tid) {
rte_call(perf, barrier);
/* Assuming all threads are fairly treated, reporting only tid==0
TODO: aggregate reports */
ucx_perf_calc_result(perf, result);
rte_call(perf, report, result, 1);
}
out:
return &statuses[tid];
}