Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
Added support for sampling up to a given number of iterations or until a threshold...
[simgrid.git] / src / smpi / smpi_bench.c
index 81ed0f0..084123b 100644 (file)
-#include <stdio.h>
-#include "private.h"
+/* Copyright (c) 2007, 2009, 2010. The SimGrid Team.
+ * All rights reserved.                                                     */
 
-// FIXME: could cause trouble with multithreaded procs on same host...
-// FIXME: add benchmarking flag?
+/* 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. */
 
-void smpi_bench_begin()
+#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)");
+
+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)
 {
-       SIMIX_mutex_lock(smpi_global->timer_mutex);
-
-       xbt_os_timer_start(smpi_global->timer);
-
-       return;
+  if (allocs) {
+    xbt_dict_free(&allocs);
+  }
+  if (samples) {
+    xbt_dict_free(&samples);
+  }
 }
 
-double smpi_bench_end()
+static void smpi_execute_flops(double flops)
 {
-       double duration;
-       smx_host_t host;
-       smx_action_t action;
-
-       xbt_os_timer_stop(smpi_global->timer);
+  smx_action_t action;
+  smx_host_t host;
+  host = SIMIX_host_self();
+
+  DEBUG1("Handle real computation time: %f flops", flops);
+  action = SIMIX_req_host_execute("computation", host, flops);
+#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);
+}
 
-       duration = xbt_os_timer_elapsed(smpi_global->timer);
+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);
+    smpi_execute_flops(duration *
+                       xbt_cfg_get_double(_surf_cfg_set,
+                                          "smpi/running_power"));
+  }
+}
 
-       host     = SIMIX_host_self();
-       action   = SIMIX_action_execute(host, "computation", duration * SMPI_DEFAULT_SPEED);
+void smpi_bench_begin(void)
+{
+  xbt_os_timer_start(smpi_process_timer());
+}
 
-       SIMIX_register_action_to_condition(action, smpi_global->timer_cond);
-       SIMIX_cond_wait(smpi_global->timer_cond, smpi_global->timer_mutex);
-       SIMIX_unregister_action_to_condition(action, smpi_global->timer_cond);
+void smpi_bench_end(void)
+{
+  xbt_os_timer_t timer = smpi_process_timer();
 
-       SIMIX_mutex_unlock(smpi_global->timer_mutex);
+  xbt_os_timer_stop(timer);
+  smpi_execute(xbt_os_timer_elapsed(timer));
+}
 
-       return duration;
+unsigned int smpi_sleep(unsigned int secs)
+{
+  smpi_execute((double) secs);
+  return secs;
 }
 
-void smpi_bench_skip() {
-       smx_host_t host;
-       smx_action_t action;
-       double duration = smpi_global->times[0];
+int smpi_gettimeofday(struct timeval *tv, struct timezone *tz)
+{
+  double now = SIMIX_get_clock();
 
-       SIMIX_mutex_lock(smpi_global->timer_mutex);
+  if (tv) {
+    tv->tv_sec = (time_t) now;
+    tv->tv_usec = (suseconds_t) (now * 1e6);
+  }
+  return 0;
+}
 
-       host   = SIMIX_host_self();
-        action = SIMIX_action_execute(host, "computation", duration * SMPI_DEFAULT_SPEED);
+static char *sample_location(int global, const char *file, int line)
+{
+  if (global) {
+    return bprintf("%s:%d", file, line);
+  } else {
+    return bprintf("%s:%d:%d", file, line, smpi_process_index());
+  }
+}
 
-       SIMIX_register_action_to_condition(action, smpi_global->timer_cond);
-       SIMIX_cond_wait(smpi_global->timer_cond, smpi_global->timer_mutex);
-       SIMIX_unregister_action_to_condition(action, smpi_global->timer_cond);
+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;
+
+  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);
+  }
+  free(loc);
+}
 
-       SIMIX_mutex_unlock(smpi_global->timer_mutex);
+int smpi_sample_2(int global, const char *file, int line)
+{
+  char *loc = sample_location(global, file, line);
+  local_data_t *data;
+
+  xbt_assert0(samples, "You did something very inconsistent, didn't you?");
+  data = xbt_dict_get_or_null(samples, loc);
+  if (!data) {
+    xbt_assert0(data, "Please, do thing in order");
+  }
+  if (!data->started) {
+    if ((data->iters > 0 && data->count >= data->iters)
+        || (data->count > 1 && data->threshold > 0.0 && data->relstderr <= data->threshold)) {
+      DEBUG1("Perform some wait of %f", data->mean);
+      smpi_execute(data->mean);
+    } else {
+      data->started = 1;
+      data->count++;
+    }
+  } else {
+    data->started = 0;
+  }
+  free(loc);
+  smpi_bench_begin();
+  smpi_process_simulated_start();
+  return data->started;
+}
 
-       return;
+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->iters) {
+    xbt_assert0(data, "Please, do thing in order");
+  }
+  smpi_bench_end();
+  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;
+  DEBUG4("Average mean after %d steps is %f, relative standard error is %f (sample was %f)", data->count,
+         data->mean, data->relstderr, sample);
 }
 
-void smpi_do_once_1() {
-       smpi_bench_end();
-       SIMIX_mutex_lock(smpi_global->times_mutex);
-       if (0 < smpi_global->times[0]) {
-               smpi_bench_skip();
-       }
-       return;
+void smpi_sample_flops(double flops)
+{
+  smpi_execute_flops(flops);
 }
 
-int smpi_do_once_2() {
-       int retval = 1;
-       if (0 < smpi_global->times[0]) {
-               SIMIX_mutex_unlock(smpi_global->times_mutex);
-               retval = 0;
-       }
-       smpi_bench_begin();
-       return retval;
+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() {
-       smpi_global->times[0] = smpi_bench_end();
-       return;
+void smpi_shared_free(void *ptr)
+{
+  shared_data_t *data = (shared_data_t *) ((int *) ptr - 1);
+  char *loc;
+
+  if (!allocs) {
+    WARN0("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);
+    return;
+  }
+  data->count--;
+  if (data->count <= 0) {
+    xbt_dict_remove(allocs, loc);
+  }
 }