/* 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 <math.h> // sqrt
#include "private.h"
#include "xbt/dict.h"
#include "xbt/sysdep.h"
+#include "xbt/ex.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 */
+xbt_dict_t calls = NULL; /* Allocated on first use */
typedef struct {
int count;
} 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;
void smpi_bench_destroy(void)
{
- if (allocs) {
- xbt_dict_free(&allocs);
- }
- if (samples) {
- xbt_dict_free(&samples);
- }
+ xbt_dict_free(&allocs);
+ xbt_dict_free(&samples);
+ xbt_dict_free(&calls);
}
static void smpi_execute_flops(double flops)
{
- smx_host_t host;
smx_action_t action;
- smx_mutex_t mutex;
- smx_cond_t cond;
- e_surf_action_state_t state;
-
+ smx_host_t host;
host = SIMIX_host_self();
- mutex = SIMIX_mutex_init();
- cond = SIMIX_cond_init();
- DEBUG1("Handle real computation time: %f flops", flops);
- action = SIMIX_action_execute(host, "computation", flops);
- 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_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")) {
- 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"));
}
}
-void smpi_bench_begin(int rank, const char *mpi_call)
+void smpi_bench_begin(void)
{
- if (mpi_call && rank >= 0
- && xbt_cfg_get_int(_surf_cfg_set, "smpi/log_events")) {
- INFO3("SMPE: ts=%f rank=%d type=end et=%s", SIMIX_get_clock(), rank,
- mpi_call);
- }
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, "smpi/log_events")) {
- INFO3("SMPE: ts=%f rank=%d type=begin et=%s", SIMIX_get_clock(), rank,
- mpi_call);
- }
}
unsigned int smpi_sleep(unsigned int secs)
}
}
-void smpi_sample_1(int global, const char *file, int line, int max)
+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(-1, NULL); /* Take time from previous MPI call into account */
+ 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->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);
+ 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_assert0(samples, "You did something very inconsistent, didn't you?");
+ xbt_assert(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");
+ xbt_assert(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;
}
free(loc);
- smpi_bench_begin(-1, NULL);
+ smpi_bench_begin();
smpi_process_simulated_start();
return data->started;
}
{
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?");
+ xbt_assert(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) {
- xbt_assert0(data, "Please, do thing in order");
+ 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);
}
- smpi_bench_end(-1, NULL);
- data->time += smpi_process_simulated_elapsed();
- DEBUG2("Average mean after %d steps is %f", data->count,
- data->time / (double) data->count);
+ free(loc);
}
void smpi_sample_flops(double flops)
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--;
xbt_dict_remove(allocs, loc);
}
}
+
+int smpi_shared_known_call(const char* func, const char* input) {
+ char* loc = bprintf("%s:%s", func, input);
+ xbt_ex_t ex;
+ int known;
+
+ if(!calls) {
+ calls = xbt_dict_new();
+ }
+ TRY {
+ xbt_dict_get(calls, loc); /* Succeed or throw */
+ known = 1;
+ }
+ CATCH(ex) {
+ if(ex.category == not_found_error) {
+ known = 0;
+ xbt_ex_free(ex);
+ } else {
+ RETHROW;
+ }
+ }
+ free(loc);
+ return known;
+}
+
+void* smpi_shared_get_call(const char* func, const char* input) {
+ char* loc = bprintf("%s:%s", func, input);
+ void* data;
+
+ if(!calls) {
+ calls = xbt_dict_new();
+ }
+ data = xbt_dict_get(calls, loc);
+ free(loc);
+ return data;
+}
+
+void* smpi_shared_set_call(const char* func, const char* input, void* data) {
+ char* loc = bprintf("%s:%s", func, input);
+
+ if(!calls) {
+ calls = xbt_dict_new();
+ }
+ xbt_dict_set(calls, loc, data, NULL);
+ free(loc);
+ return data;
+}
