+/* Copyright (c) 2007, 2009, 2010. The SimGrid Team.
+ * All rights reserved. */
+
+/* This program is free software; you can redistribute it and/or modify it
+ * under the terms of the license (GNU LGPL) which comes with this package. */
+
#include "private.h"
+#include "xbt/dict.h"
+#include "xbt/sysdep.h"
+#include "surf/surf.h"
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi,
"Logging specific to SMPI (benchmarking)");
-static void smpi_execute(double duration) {
- smx_host_t host;
- smx_action_t action;
- smx_mutex_t mutex;
- smx_cond_t cond;
- e_surf_action_state_t state;
-
- if(duration > 0.001) {
- host = SIMIX_host_self();
- mutex = SIMIX_mutex_init();
- cond = SIMIX_cond_init();
- DEBUG1("Sleep for %f to handle real computation time", duration);
- duration *= xbt_cfg_get_double(_surf_cfg_set, "reference_speed");
- action = SIMIX_action_execute(host, "computation", duration);
- SIMIX_mutex_lock(mutex);
- SIMIX_register_action_to_condition(action, cond);
- for(state = SIMIX_action_get_state(action);
- state == SURF_ACTION_READY ||
- state == SURF_ACTION_RUNNING; state = SIMIX_action_get_state(action)) {
- SIMIX_cond_wait(cond, mutex);
- }
- SIMIX_unregister_action_to_condition(action, cond);
- SIMIX_mutex_unlock(mutex);
- SIMIX_action_destroy(action);
- SIMIX_cond_destroy(cond);
- SIMIX_mutex_destroy(mutex);
+xbt_dict_t allocs = NULL; /* Allocated on first use */
+xbt_dict_t samples = NULL; /* Allocated on first use */
+
+typedef struct {
+ int count;
+ char data[];
+} shared_data_t;
+
+typedef struct {
+ int count;
+ double sum;
+ double sum_pow2;
+ double mean;
+ double relstderr;
+ int iters;
+ double threshold;
+ int started;
+} local_data_t;
+
+void smpi_bench_destroy(void)
+{
+ if (allocs) {
+ xbt_dict_free(&allocs);
+ }
+ if (samples) {
+ xbt_dict_free(&samples);
}
}
-void smpi_bench_begin(int rank, const char* mpi_call) {
- if(mpi_call && rank >= 0 && xbt_cfg_get_int(_surf_cfg_set, "SMPE")) {
- INFO3("SMPE: ts=%f rank=%d type=end et=%s", SIMIX_get_clock(), rank, mpi_call);
+static void smpi_execute_flops(double flops)
+{
+ smx_action_t action;
+ smx_host_t host;
+ host = SIMIX_host_self();
+
+ 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);
+}
+
+static void smpi_execute(double duration)
+{
+ if (duration >= xbt_cfg_get_double(_surf_cfg_set, "smpi/cpu_threshold")) {
+ 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"));
}
+}
+
+void smpi_bench_begin(void)
+{
xbt_os_timer_start(smpi_process_timer());
}
-void smpi_bench_end(int rank, const char* mpi_call) {
+void smpi_bench_end(void)
+{
xbt_os_timer_t timer = smpi_process_timer();
xbt_os_timer_stop(timer);
smpi_execute(xbt_os_timer_elapsed(timer));
- if(mpi_call && rank >= 0 && xbt_cfg_get_int(_surf_cfg_set, "SMPE")) {
- INFO3("SMPE: ts=%f rank=%d type=begin et=%s", SIMIX_get_clock(), rank, mpi_call);
+}
+
+unsigned int smpi_sleep(unsigned int secs)
+{
+ smpi_execute((double) secs);
+ return secs;
+}
+
+int smpi_gettimeofday(struct timeval *tv, struct timezone *tz)
+{
+ double now = SIMIX_get_clock();
+
+ if (tv) {
+ tv->tv_sec = (time_t) now;
+ tv->tv_usec = (suseconds_t) (now * 1e6);
}
+ return 0;
}
-/*
-TODO
-void smpi_do_once_1(const char *file, int line)
+static char *sample_location(int global, const char *file, int line)
{
- smpi_do_once_duration_node_t curr, prev;
+ if (global) {
+ return bprintf("%s:%d", file, line);
+ } else {
+ return bprintf("%s:%d:%d", file, line, smpi_process_index());
+ }
+}
- smpi_bench_end();
- SIMIX_mutex_lock(smpi_global->do_once_mutex);
- prev = NULL;
- for(curr = smpi_global->do_once_duration_nodes;
- NULL != curr && (strcmp(curr->file, file) || curr->line != line);
- curr = curr->next) {
- prev = curr;
+int 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;
+
+ smpi_bench_end(); /* Take time from previous MPI call into account */
+ if (!samples) {
+ samples = xbt_dict_new();
+ }
+ data = xbt_dict_get_or_null(samples, loc);
+ if (!data) {
+ data = (local_data_t *) xbt_new(local_data_t, 1);
+ data->count = 0;
+ 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);
+ return 0;
+ }
+ free(loc);
+ return 1;
+}
+
+int smpi_sample_2(int global, const char *file, int line)
+{
+ char *loc = sample_location(global, file, line);
+ local_data_t *data;
+
+ xbt_assert(samples, "You did something very inconsistent, didn't you?");
+ data = xbt_dict_get_or_null(samples, loc);
+ if (!data) {
+ xbt_assert(data, "Please, do thing in order");
}
- if(NULL == curr) {
- curr = xbt_new(s_smpi_do_once_duration_node_t, 1);
- curr->file = xbt_strdup(file);
- curr->line = line;
- curr->duration = -1;
- curr->next = NULL;
- if(NULL == prev) {
- smpi_global->do_once_duration_nodes = curr;
+ if (!data->started) {
+ 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 {
- prev->next = curr;
+ data->started = 1;
+ data->count++;
}
+ } else {
+ data->started = 0;
}
- smpi_global->do_once_duration = &curr->duration;
+ free(loc);
+ smpi_bench_begin();
+ smpi_process_simulated_start();
+ return data->started;
}
-int smpi_do_once_2()
+void smpi_sample_3(int global, const char *file, int line)
{
- double duration = *(smpi_global->do_once_duration);
+ char *loc = sample_location(global, file, line);
+ local_data_t *data;
+ double sample, n;
- if(0 > duration) {
- smpi_start_timer();
- return 1;
+ xbt_assert(samples, "You did something very inconsistent, didn't you?");
+ data = xbt_dict_get_or_null(samples, loc);
+ smpi_bench_end();
+ if(data && data->started && data->count < data->iters) {
+ 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);
}
- SIMIX_mutex_unlock(smpi_global->do_once_mutex);
- smpi_execute(duration);
- smpi_bench_begin();
- return 0;
+ free(loc);
+}
+
+void smpi_sample_flops(double flops)
+{
+ smpi_execute_flops(flops);
+}
+
+void *smpi_shared_malloc(size_t size, const char *file, int line)
+{
+ char *loc = bprintf("%s:%d:%zu", file, line, size);
+ shared_data_t *data;
+
+ if (!allocs) {
+ allocs = xbt_dict_new();
+ }
+ data = xbt_dict_get_or_null(allocs, loc);
+ if (!data) {
+ data = (shared_data_t *) xbt_malloc0(sizeof(int) + size);
+ data->count = 1;
+ xbt_dict_set(allocs, loc, data, &free);
+ } else {
+ data->count++;
+ }
+ free(loc);
+ return data->data;
}
-void smpi_do_once_3()
+void smpi_shared_free(void *ptr)
{
- *(smpi_global->do_once_duration) = smpi_stop_timer();
+ shared_data_t *data = (shared_data_t *) ((int *) ptr - 1);
+ char *loc;
+
+ if (!allocs) {
+ XBT_WARN("Cannot free: nothing was allocated");
+ return;
+ }
+ loc = xbt_dict_get_key(allocs, data);
+ if (!loc) {
+ XBT_WARN("Cannot free: %p was not shared-allocated by SMPI", ptr);
+ return;
+ }
+ data->count--;
+ if (data->count <= 0) {
+ xbt_dict_remove(allocs, loc);
+ }
}
-*/