#include "smpi_comm.hpp"
#include "src/internal_config.h"
#include "src/mc/mc_replay.hpp"
-#include "src/simix/ActorImpl.hpp"
#include "xbt/config.hpp"
#include "src/smpi/include/smpi_actor.hpp"
SharedMallocType smpi_cfg_shared_malloc = SharedMallocType::GLOBAL;
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);
-}
-
-extern "C" XBT_PUBLIC void smpi_execute_(double* duration);
-void smpi_execute_(double *duration)
-{
- smpi_execute(*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);
simgrid::s4u::this_actor::exec_init(flops)
->set_name("computation")
- ->set_tracing_category(TRACE_internal_smpi_get_category())
+ ->set_tracing_category(smpi_process()->get_tracing_category())
->start()
->wait();
smpi_switch_data_segment(simgrid::s4u::Actor::self());
{
if (not smpi_process())
return sleep(secs);
- return private_sleep(static_cast<double>(secs));
+ return private_sleep(secs);
}
int smpi_usleep(useconds_t usecs)
{
if (not smpi_process())
return usleep(usecs);
- return static_cast<int>(private_sleep(static_cast<double>(usecs) / 1000000.0));
+ return static_cast<int>(private_sleep(usecs / 1000000.0));
}
#if _POSIX_TIMERS > 0
{
if (not smpi_process())
return nanosleep(tp,t);
- return static_cast<int>(private_sleep(static_cast<double>(tp->tv_sec + tp->tv_nsec / 1000000000.0)));
+ return static_cast<int>(private_sleep(tp->tv_sec + tp->tv_nsec / 1000000000.0));
}
#endif
}
}
-int smpi_sample_2(int global, const char *file, int line)
+int smpi_sample_2(int global, const char *file, int line, int iter_count)
{
SampleLocation loc(global, file, line);
- int res;
- XBT_DEBUG("sample2 %s", loc.c_str());
+ XBT_DEBUG("sample2 %s %d", loc.c_str(), iter_count);
auto sample = samples.find(loc);
if (sample == samples.end())
xbt_die("Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f; total:%f",
data.count, data.iters, data.relstderr, data.threshold, data.mean, data.sum);
smpi_bench_begin();
- res = 1;
} 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
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). "
- "Mean is %f %s",
- data.count, data.iters, data.relstderr, data.threshold, data.mean,
- (data.sum != 0.0 ? "Applying it count times because we finished benching count iterations": ""));
- double sleep = data.mean;
- if (data.sum != 0.0){ //we finished benching, sum is unecessary after the first injection, we can reset it.
- sleep = data.sum;
- data.sum = 0.0;
- }
+ "Mean is %f, will be injected %d times",
+ data.count, data.iters, data.relstderr, data.threshold, data.mean, iter_count);
+
+ //we ended benchmarking, let's inject all the time, now, and fast forward out of the loop.
smpi_process()->set_sampling(0);
- smpi_execute(sleep);
+ smpi_execute(data.mean*iter_count);
smpi_bench_begin();
+ return 0;
} else {
XBT_DEBUG("Skipping - Benchmark already performed - accumulating time");
xbt_os_threadtimer_start(smpi_process()->timer());
}
- res = 0; // prepare to capture future, unrelated computations
}
-
- return res;
+ return 1;
}
void smpi_sample_3(int global, const char *file, int line)
double period = xbt_os_timer_elapsed(smpi_process()->timer());
data.sum += period;
data.sum_pow2 += period * period;
- double n = static_cast<double>(data.count);
+ double n = data.count;
data.mean = data.sum / n;
data.relstderr = sqrt((data.sum_pow2 / n - data.mean * data.mean) / n) / data.mean;
data.benching = false;
}
-void smpi_sample_exit(int global, const char *file, int line){
+int smpi_sample_exit(int global, const char *file, int line, int iter_count){
if (smpi_process()->sampling()){
SampleLocation loc(global, file, line);
LocalData& data = sample->second;
if (smpi_process()->sampling()){//end of loop, but still sampling needed
- double sleep = data.mean;
- if (data.sum != 0.0){ //we finished benching, sum is unecessary after the first injection, we can reset it.
- sleep = data.sum;
- data.sum = 0.0;
- }
smpi_process()->set_sampling(0);
- smpi_execute(sleep);
+ smpi_execute(data.mean*iter_count);
smpi_bench_begin();
}
}
+ return 0;
}
smpi_trace_call_location_t* smpi_trace_get_call_location()