X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/005cf0f71b3817a9e93de0b0c3bd92f216ef5c7f..6ee7e9c2e455536ab817ae0136acfbb53822eecd:/src/smpi/smpi_bench.c diff --git a/src/smpi/smpi_bench.c b/src/smpi/smpi_bench.c index 4ecf4ebbb4..d5eb6b3e9f 100644 --- a/src/smpi/smpi_bench.c +++ b/src/smpi/smpi_bench.c @@ -21,9 +21,13 @@ typedef struct { } shared_data_t; typedef struct { - double time; int count; - int max; + double sum; + double sum_pow2; + double mean; + double relstderr; + int iters; + double threshold; int started; } local_data_t; @@ -43,16 +47,18 @@ static void smpi_execute_flops(double flops) smx_host_t host; host = SIMIX_host_self(); - DEBUG1("Handle real computation time: %f flops", flops); - action = SIMIX_req_host_execute("computation", host, flops); + XBT_DEBUG("Handle real computation time: %f flops", flops); + action = SIMIX_req_host_execute("computation", host, flops, 1); +#ifdef HAVE_TRACING + SIMIX_req_set_category (action, TRACE_internal_smpi_get_category()); +#endif SIMIX_req_host_execution_wait(action); - SIMIX_req_host_execution_destroy(action); } static void smpi_execute(double duration) { if (duration >= xbt_cfg_get_double(_surf_cfg_set, "smpi/cpu_threshold")) { - DEBUG1("Sleep for %f to handle real computation time", duration); + XBT_DEBUG("Sleep for %f to handle real computation time", duration); smpi_execute_flops(duration * xbt_cfg_get_double(_surf_cfg_set, "smpi/running_power")); @@ -98,7 +104,7 @@ static char *sample_location(int global, const char *file, int line) } } -void smpi_sample_1(int global, const char *file, int line, int max) +void smpi_sample_1(int global, const char *file, int line, int iters, double threshold) { char *loc = sample_location(global, file, line); local_data_t *data; @@ -110,9 +116,11 @@ void smpi_sample_1(int global, const char *file, int line, int max) data = xbt_dict_get_or_null(samples, loc); if (!data) { data = (local_data_t *) xbt_new(local_data_t, 1); - data->time = 0.0; data->count = 0; - data->max = max; + data->sum = 0.0; + data->sum_pow2 = 0.0; + data->iters = iters; + data->threshold = threshold; data->started = 0; xbt_dict_set(samples, loc, data, &free); } @@ -130,12 +138,13 @@ int smpi_sample_2(int global, const char *file, int line) xbt_assert0(data, "Please, do thing in order"); } if (!data->started) { - if (data->count < data->max) { + if ((data->iters > 0 && data->count >= data->iters) + || (data->count > 1 && data->threshold > 0.0 && data->relstderr <= data->threshold)) { + XBT_DEBUG("Perform some wait of %f", data->mean); + smpi_execute(data->mean); + } else { data->started = 1; data->count++; - } else { - DEBUG1("Perform some wait of %f", data->time / (double) data->count); - smpi_execute(data->time / (double) data->count); } } else { data->started = 0; @@ -150,16 +159,22 @@ void smpi_sample_3(int global, const char *file, int line) { char *loc = sample_location(global, file, line); local_data_t *data; + double sample, n; xbt_assert0(samples, "You did something very inconsistent, didn't you?"); data = xbt_dict_get_or_null(samples, loc); - if (!data || !data->started || data->count >= data->max) { + if (!data || !data->started || data->count >= data->iters) { xbt_assert0(data, "Please, do thing in order"); } smpi_bench_end(); - data->time += smpi_process_simulated_elapsed(); - DEBUG2("Average mean after %d steps is %f", data->count, - data->time / (double) data->count); + sample = smpi_process_simulated_elapsed(); + data->sum += sample; + data->sum_pow2 += sample * sample; + n = (double)data->count; + data->mean = data->sum / n; + data->relstderr = sqrt((data->sum_pow2 / n - data->mean * data->mean) / n) / data->mean; + XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)", data->count, + data->mean, data->relstderr, sample); } void smpi_sample_flops(double flops) @@ -193,12 +208,12 @@ void smpi_shared_free(void *ptr) char *loc; if (!allocs) { - WARN0("Cannot free: nothing was allocated"); + XBT_WARN("Cannot free: nothing was allocated"); return; } loc = xbt_dict_get_key(allocs, data); if (!loc) { - WARN1("Cannot free: %p was not shared-allocated by SMPI", ptr); + XBT_WARN("Cannot free: %p was not shared-allocated by SMPI", ptr); return; } data->count--;