A
lgorithmique
N
umérique
D
istribuée
Public GIT Repository
projects
/
simgrid.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
sanitize the OOP of kernel::profile
[simgrid.git]
/
src
/
smpi
/
internals
/
smpi_bench.cpp
diff --git
a/src/smpi/internals/smpi_bench.cpp
b/src/smpi/internals/smpi_bench.cpp
index
bc55483
..
9fbc6b0
100644
(file)
--- a/
src/smpi/internals/smpi_bench.cpp
+++ b/
src/smpi/internals/smpi_bench.cpp
@@
-1,4
+1,4
@@
-/* Copyright (c) 2007-201
8
. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2007-201
9
. 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. */
/* 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. */
@@
-28,6
+28,11
@@
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)");
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)");
+static simgrid::config::Flag<double>
+ smpi_wtime_sleep("smpi/wtime",
+ "Minimum time to inject inside a call to MPI_Wtime(), gettimeofday() and clock_gettime()",
+ 1e-8 /* Documented to be 10 ns */);
+
double smpi_cpu_threshold = -1;
double smpi_host_speed;
double smpi_cpu_threshold = -1;
double smpi_host_speed;
@@
-35,6
+40,7
@@
SharedMallocType smpi_cfg_shared_malloc = SharedMallocType::GLOBAL;
double smpi_total_benched_time = 0;
extern "C" XBT_PUBLIC void smpi_execute_flops_(double* flops);
double smpi_total_benched_time = 0;
extern "C" XBT_PUBLIC void smpi_execute_flops_(double* flops);
+
void smpi_execute_flops_(double *flops)
{
smpi_execute_flops(*flops);
void smpi_execute_flops_(double *flops)
{
smpi_execute_flops(*flops);
@@
-49,10
+55,11
@@
void smpi_execute_(double *duration)
void smpi_execute_flops(double flops) {
xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
XBT_DEBUG("Handle real computation time: %f flops", flops);
void smpi_execute_flops(double flops) {
xbt_assert(flops >= 0, "You're trying to execute a negative amount of flops (%f)!", flops);
XBT_DEBUG("Handle real computation time: %f flops", flops);
- simgrid::s4u::ExecPtr e = simgrid::s4u::this_actor::exec_init(flops)->set_name("computation");
- e->start();
- e->set_tracing_category(TRACE_internal_smpi_get_category());
- e->wait();
+ simgrid::s4u::this_actor::exec_init(flops)
+ ->set_name("computation")
+ ->set_tracing_category(TRACE_internal_smpi_get_category())
+ ->start()
+ ->wait();
smpi_switch_data_segment(simgrid::s4u::Actor::self());
}
smpi_switch_data_segment(simgrid::s4u::Actor::self());
}
@@
-95,12
+102,10
@@
void smpi_bench_begin()
if (not simgrid::config::get_value<std::string>("smpi/papi-events").empty()) {
int event_set = smpi_process()->papi_event_set();
// PAPI_start sets everything to 0! See man(3) PAPI_start
if (not simgrid::config::get_value<std::string>("smpi/papi-events").empty()) {
int event_set = smpi_process()->papi_event_set();
// PAPI_start sets everything to 0! See man(3) PAPI_start
- if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized()) {
- if (PAPI_start(event_set) != PAPI_OK) {
- // TODO This needs some proper handling.
- XBT_CRITICAL("Could not start PAPI counters.\n");
- xbt_die("Error.");
- }
+ if (PAPI_LOW_LEVEL_INITED == PAPI_is_initialized() && PAPI_start(event_set) != PAPI_OK) {
+ // TODO This needs some proper handling.
+ XBT_CRITICAL("Could not start PAPI counters.\n");
+ xbt_die("Error.");
}
}
#endif
}
}
#endif
@@
-176,7
+181,7
@@
void smpi_bench_end()
smpi_total_benched_time += xbt_os_timer_elapsed(timer);
}
smpi_total_benched_time += xbt_os_timer_elapsed(timer);
}
-/* Private sleep function used by smpi_sleep()
and smpi_usleep()
*/
+/* Private sleep function used by smpi_sleep()
, smpi_usleep() and friends
*/
static unsigned int private_sleep(double secs)
{
smpi_bench_end();
static unsigned int private_sleep(double secs)
{
smpi_bench_end();
@@
-231,6
+236,8
@@
int smpi_gettimeofday(struct timeval* tv, struct timezone* tz)
tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
#endif
}
tv->tv_usec = static_cast<suseconds_t>((now - tv->tv_sec) * 1e6);
#endif
}
+ if (smpi_wtime_sleep > 0)
+ simcall_process_sleep(smpi_wtime_sleep);
smpi_bench_begin();
return 0;
}
smpi_bench_begin();
return 0;
}
@@
-247,11
+254,28
@@
int smpi_clock_gettime(clockid_t clk_id, struct timespec* tp)
tp->tv_sec = static_cast<time_t>(now);
tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
}
tp->tv_sec = static_cast<time_t>(now);
tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
}
+ if (smpi_wtime_sleep > 0)
+ simcall_process_sleep(smpi_wtime_sleep);
smpi_bench_begin();
return 0;
}
#endif
smpi_bench_begin();
return 0;
}
#endif
+double smpi_mpi_wtime()
+{
+ double time;
+ if (smpi_process()->initialized() && not smpi_process()->finalized() && not smpi_process()->sampling()) {
+ smpi_bench_end();
+ time = SIMIX_get_clock();
+ if (smpi_wtime_sleep > 0)
+ simcall_process_sleep(smpi_wtime_sleep);
+ smpi_bench_begin();
+ } else {
+ time = SIMIX_get_clock();
+ }
+ return time;
+}
+
extern double sg_surf_precision;
unsigned long long smpi_rastro_resolution ()
{
extern double sg_surf_precision;
unsigned long long smpi_rastro_resolution ()
{
@@
-367,9
+391,12
@@
int smpi_sample_2(int global, const char *file, int line)
} else {
// Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
//ran one bench and need to bail out now that our job is done). Just sleep instead
} else {
// Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just
//ran one bench and need to bail out now that our job is done). Just sleep instead
- XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f)."
+ if (not data.need_more_benchs())
+ XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f)."
" apply the %fs delay instead",
data.count, data.iters, data.relstderr, data.threshold, data.mean);
" apply the %fs delay instead",
data.count, data.iters, data.relstderr, data.threshold, data.mean);
+ else
+ XBT_DEBUG("Skipping - Benchmark already performed");
smpi_execute(data.mean);
smpi_process()->set_sampling(0);
res = 0; // prepare to capture future, unrelated computations
smpi_execute(data.mean);
smpi_process()->set_sampling(0);
res = 0; // prepare to capture future, unrelated computations