#include "src/internal_config.h"
#include "private.h"
#include "private.hpp"
+#include <xbt/ex.hpp>
#include "xbt/dict.h"
#include "xbt/sysdep.h"
#include "xbt/ex.h"
#include <string.h>
#include <stdio.h>
+#if HAVE_PAPI
+#include <papi.h>
+#endif
+
#ifndef MAP_ANONYMOUS
#define MAP_ANONYMOUS MAP_ANON
#endif
class smpi_source_location {
public:
smpi_source_location(const char* filename, int line)
- : filename(filename), filename_length(strlen(filename)), line(line) {}
+ : filename(xbt_strdup(filename)), filename_length(strlen(filename)), line(line) {}
/** Pointer to a static string containing the file name */
- const char* filename = nullptr;
+ char* filename = nullptr;
int filename_length = 0;
int line = 0;
mem = mmap(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
if(mem == MAP_FAILED) {
- xbt_die("Could not map fd %d: %s", fd, strerror(errno));
+ xbt_die("Could not map fd %d with size %zu: %s", fd, size, strerror(errno));
}
snprintf(loc, PTR_STRLEN, "%p", mem);
meta.size = size;
}
#endif
-void smpi_bench_destroy(void)
+void smpi_bench_destroy()
{
allocs.clear();
allocs_metadata.clear();
}
void smpi_execute_flops(double flops) {
- smx_synchro_t action;
+ smx_activity_t action;
XBT_DEBUG("Handle real computation time: %f flops", flops);
- action = simcall_execution_start("computation", flops, 1, 0, 0);
+ action = simcall_execution_start("computation", flops, 1, 0);
simcall_set_category (action, TRACE_internal_smpi_get_category());
simcall_execution_wait(action);
smpi_switch_data_segment(smpi_process_index());
}
}
-void smpi_switch_data_segment(int dest);
-
-void smpi_bench_begin(void)
+void smpi_bench_begin()
{
if (smpi_privatize_global_variables) {
smpi_switch_data_segment(smpi_process_index());
if (MC_is_active() || MC_record_replay_is_active())
return;
+#if HAVE_PAPI
+ if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
+ 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.");
+ }
+ }
+ }
+#endif
xbt_os_threadtimer_start(smpi_process_timer());
}
-void smpi_bench_end(void)
+void smpi_bench_end()
{
-
if (MC_is_active() || MC_record_replay_is_active())
return;
double speedup = 1;
xbt_os_timer_t timer = smpi_process_timer();
xbt_os_threadtimer_stop(timer);
+
+#if HAVE_PAPI
+ /**
+ * An MPI function has been called and now is the right time to update
+ * our PAPI counters for this process.
+ */
+ if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
+ papi_counter_t& counter_data = smpi_process_papi_counters();
+ int event_set = smpi_process_papi_event_set();
+ std::vector<long long> event_values = std::vector<long long>(counter_data.size());
+
+ if (PAPI_stop(event_set, &event_values[0]) != PAPI_OK) { // Error
+ XBT_CRITICAL("Could not stop PAPI counters.\n");
+ xbt_die("Error.");
+ } else {
+ for (unsigned int i = 0; i < counter_data.size(); i++) {
+ counter_data[i].second += event_values[i];
+ // XBT_DEBUG("[%i] PAPI: Counter %s: Value is now %lli (got increment by %lli\n", smpi_process_index(),
+ // counter_data[i].first.c_str(), counter_data[i].second, event_values[i]);
+ }
+ }
+ }
+#endif
+
if (smpi_process_get_sampling()) {
XBT_CRITICAL("Cannot do recursive benchmarks.");
XBT_CRITICAL("Are you trying to make a call to MPI within a SMPI_SAMPLE_ block?");
smpi_execute(xbt_os_timer_elapsed(timer)/speedup);
}
+#if HAVE_PAPI
+ if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0' && TRACE_smpi_is_enabled()) {
+ char container_name[INSTR_DEFAULT_STR_SIZE];
+ smpi_container(smpi_process_index(), container_name, INSTR_DEFAULT_STR_SIZE);
+ container_t container = PJ_container_get(container_name);
+ papi_counter_t& counter_data = smpi_process_papi_counters();
+
+ for (auto& pair : counter_data) {
+ new_pajeSetVariable(surf_get_clock(), container,
+ PJ_type_get(/* countername */ pair.first.c_str(), container->type), pair.second);
+ }
+ }
+#endif
+
smpi_total_benched_time += xbt_os_timer_elapsed(timer);
}
return static_cast<int>(private_sleep(static_cast<double>(usecs) / 1000000.0));
}
+#if _POSIX_TIMERS > 0
+int smpi_nanosleep(const struct timespec *tp, struct timespec * t)
+{
+ return static_cast<int>(private_sleep(static_cast<double>(tp->tv_sec + tp->tv_nsec / 1000000000.0)));
+}
+#endif
+
int smpi_gettimeofday(struct timeval *tv, void* tz)
{
double now;
return 0;
}
+#if _POSIX_TIMERS > 0
+int smpi_clock_gettime(clockid_t clk_id, struct timespec *tp)
+{
+ //there is only one time in SMPI, so clk_id is ignored.
+ double now;
+ smpi_bench_end();
+ now = SIMIX_get_clock();
+ if (tp) {
+ tp->tv_sec = static_cast<time_t>(now);
+ tp->tv_nsec = static_cast<long int>((now - tp->tv_sec) * 1e9);
+ }
+ smpi_bench_begin();
+ return 0;
+}
+#endif
+
extern double sg_surf_precision;
-unsigned long long smpi_rastro_resolution (void)
+unsigned long long smpi_rastro_resolution ()
{
smpi_bench_end();
double resolution = (1/sg_surf_precision);
return static_cast<unsigned long long>(resolution);
}
-unsigned long long smpi_rastro_timestamp (void)
+unsigned long long smpi_rastro_timestamp ()
{
smpi_bench_end();
double now = SIMIX_get_clock();
void *smpi_shared_malloc(size_t size, const char *file, int line)
{
void* mem;
- if (xbt_cfg_get_boolean("smpi/use-shared-malloc")){
+ if (size > 0 && xbt_cfg_get_boolean("smpi/use-shared-malloc")){
int fd;
smpi_source_location loc(file, line);
auto res = allocs.insert(std::make_pair(loc, shared_data_t()));
snprintf(loc, PTR_STRLEN, "%p", ptr);
auto meta = allocs_metadata.find(ptr);
if (meta == allocs_metadata.end()) {
- XBT_WARN("Cannot free: %p was not shared-allocated by SMPI", ptr);
+ XBT_WARN("Cannot free: %p was not shared-allocated by SMPI - maybe its size was 0?", ptr);
return;
}
shared_data_t* data = &meta->second.data->second;
XBT_WARN("Unmapping of fd %d failed: %s", data->fd, strerror(errno));
}
data->count--;
- XBT_DEBUG("Shared free - no removal - of %p, count = %d", ptr, data->count);
if (data->count <= 0) {
close(data->fd);
allocs.erase(allocs.find(meta->second.data->first));
XBT_DEBUG("Shared free - with removal - of %p", ptr);
+ }else{
+ XBT_DEBUG("Shared free - no removal - of %p, count = %d", ptr, data->count);
}
}else{
XBT_DEBUG("Classic free of %p", ptr);