#include "simgrid/sg_config.h"
#include "src/mc/mc_replay.h"
#include "src/msg/msg_private.h"
-#include "src/simix/SynchroComm.hpp"
+#include "src/kernel/activity/SynchroComm.hpp"
-
-#include <float.h> /* DBL_MAX */
+#include <float.h> /* DBL_MAX */
+#include <fstream>
+#include <map>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
-#include <fstream>
+#include <string>
+#include <vector>
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)");
#include <boost/tokenizer.hpp>
#include <boost/algorithm/string.hpp> /* trim_right / trim_left */
+#if HAVE_PAPI
+#include "papi.h"
+const char* papi_default_config_name = "default";
+
+struct papi_process_data {
+ papi_counter_t counter_data;
+ int event_set;
+};
+
+#endif
std::unordered_map<std::string, double> location2speedup;
typedef struct s_smpi_process_data {
int sampling; /* inside an SMPI_SAMPLE_ block? */
char* instance_id;
bool replaying; /* is the process replaying a trace */
- xbt_bar_t finalization_barrier;
+ msg_bar_t finalization_barrier;
int return_value;
- smpi_trace_call_location_t* trace_call_loc;
+ smpi_trace_call_location_t trace_call_loc;
+#if HAVE_PAPI
+ /** Contains hardware data as read by PAPI **/
+ int papi_event_set;
+ papi_counter_t papi_counter_data;
+#endif
} s_smpi_process_data_t;
-static smpi_process_data_t *process_data = NULL;
+static smpi_process_data_t *process_data = nullptr;
int process_count = 0;
int smpi_universe_size = 0;
-int* index_to_process_data = NULL;
+int* index_to_process_data = nullptr;
extern double smpi_total_benched_time;
+extern xbt_dict_t smpi_type_keyvals;
+extern xbt_dict_t smpi_comm_keyvals;
xbt_os_timer_t global_timer;
MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
-MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
-MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
-MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
+MPI_Errhandler *MPI_ERRORS_RETURN = nullptr;
+MPI_Errhandler *MPI_ERRORS_ARE_FATAL = nullptr;
+MPI_Errhandler *MPI_ERRHANDLER_NULL = nullptr;
#define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
void smpi_process_init(int *argc, char ***argv)
{
- int index=-1;
- smpi_process_data_t data;
- smx_process_t proc;
- if (argc != NULL && argv != NULL) {
- proc = SIMIX_process_self();
- SIMIX_process_set_cleanup_function(proc, MSG_process_cleanup_from_SIMIX);
+ if (process_data == nullptr){
+ printf("SimGrid was not initialized properly before entering MPI_Init. Aborting, please check compilation process and use smpirun\n");
+ exit(1);
+ }
+ if (argc != nullptr && argv != nullptr) {
+ smx_actor_t proc = SIMIX_process_self();
+ proc->context->set_cleanup(&MSG_process_cleanup_from_SIMIX);
char* instance_id = (*argv)[1];
int rank = xbt_str_parse_int((*argv)[2], "Invalid rank: %s");
- index = smpi_process_index_of_smx_process(proc);
+ int index = smpi_process_index_of_smx_process(proc);
- if(index_to_process_data == NULL){
+ if(index_to_process_data == nullptr){
index_to_process_data=static_cast<int*>(xbt_malloc(SIMIX_process_count()*sizeof(int)));
}
}
MPI_Comm* temp_comm_world;
- xbt_bar_t temp_bar;
+ msg_bar_t temp_bar;
smpi_deployment_register_process(instance_id, rank, index, &temp_comm_world, &temp_bar);
- data = smpi_process_remote_data(index);
- data->comm_world = temp_comm_world;
- if(temp_bar != NULL)
+ smpi_process_data_t data = smpi_process_remote_data(index);
+ data->comm_world = temp_comm_world;
+ if(temp_bar != nullptr)
data->finalization_barrier = temp_bar;
data->index = index;
data->instance_id = instance_id;
simdata->data = data;
if (*argc > 3) {
- free((*argv)[1]);
memmove(&(*argv)[0], &(*argv)[2], sizeof(char *) * (*argc - 2));
- (*argv)[(*argc) - 1] = NULL;
- (*argv)[(*argc) - 2] = NULL;
+ (*argv)[(*argc) - 1] = nullptr;
+ (*argv)[(*argc) - 2] = nullptr;
}
(*argc)-=2;
data->argc = argc;
XBT_DEBUG("<%d> New process in the game: %p", index, proc);
}
xbt_assert(smpi_process_data(),
- "smpi_process_data() returned NULL. You probably gave a NULL parameter to MPI_Init. Although it's required by "
- "MPI-2, this is currently not supported by SMPI.");
+ "smpi_process_data() returned nullptr. You probably gave a nullptr parameter to MPI_Init. "
+ "Although it's required by MPI-2, this is currently not supported by SMPI.");
}
void smpi_process_destroy()
int index = smpi_process_index();
// wait for all pending asynchronous comms to finish
- xbt_barrier_wait(process_data[index_to_process_data[index]]->finalization_barrier);
+ MSG_barrier_wait(process_data[index_to_process_data[index]]->finalization_barrier);
}
/** @brief Check if a process is finalized */
/** @brief Check if a process is initialized */
int smpi_process_initialized()
{
- if (index_to_process_data == NULL){
+ if (index_to_process_data == nullptr){
return false;
} else{
int index = smpi_process_index();
smpi_trace_call_location_t* smpi_process_get_call_location()
{
smpi_process_data_t process_data = smpi_process_data();
- return process_data->trace_call_loc;
+ return &process_data->trace_call_loc;
}
int smpi_process_index()
{
smpi_process_data_t data = smpi_process_data();
//return -1 if not initialized
- return data != NULL ? data->index : MPI_UNDEFINED;
+ return data != nullptr ? data->index : MPI_UNDEFINED;
}
MPI_Comm smpi_process_comm_world()
{
smpi_process_data_t data = smpi_process_data();
//return MPI_COMM_NULL if not initialized
- return data != NULL ? *data->comm_world : MPI_COMM_NULL;
+ return data != nullptr ? *data->comm_world : MPI_COMM_NULL;
}
smx_mailbox_t smpi_process_mailbox()
return data->mailboxes_mutex;
}
+#if HAVE_PAPI
+int smpi_process_papi_event_set(void)
+{
+ smpi_process_data_t data = smpi_process_data();
+ return data->papi_event_set;
+}
+
+papi_counter_t& smpi_process_papi_counters(void)
+{
+ smpi_process_data_t data = smpi_process_data();
+ return data->papi_counter_data;
+}
+#endif
+
xbt_os_timer_t smpi_process_timer()
{
smpi_process_data_t data = smpi_process_data();
smpi_process_data_t data = smpi_process_data();
if(data->comm_self==MPI_COMM_NULL){
MPI_Group group = smpi_group_new(1);
- data->comm_self = smpi_comm_new(group, NULL);
+ data->comm_self = smpi_comm_new(group, nullptr);
smpi_group_set_mapping(group, smpi_process_index(), 0);
}
message, request, request->buf, request->size, request->src, request->dst, request->tag, request->flags);
}
-void smpi_comm_copy_buffer_callback(smx_synchro_t synchro, void *buff, size_t buff_size)
+void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size)
{
XBT_DEBUG("Copy the data over");
void* tmpbuff=buff;
- simgrid::simix::Comm *comm = dynamic_cast<simgrid::simix::Comm*>(synchro);
+ simgrid::kernel::activity::Comm *comm = dynamic_cast<simgrid::kernel::activity::Comm*>(synchro);
if((smpi_privatize_global_variables) && (static_cast<char*>(buff) >= smpi_start_data_exe)
&& (static_cast<char*>(buff) < smpi_start_data_exe + smpi_size_data_exe )
xbt_free(buff);
//It seems that the request is used after the call there this should be free somewhere else but where???
//xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
- comm->src_buff = NULL;
+ comm->src_buff = nullptr;
}
if(tmpbuff!=buff)xbt_free(tmpbuff);
}
-void smpi_comm_null_copy_buffer_callback(smx_synchro_t comm, void *buff, size_t buff_size)
+void smpi_comm_null_copy_buffer_callback(smx_activity_t comm, void *buff, size_t buff_size)
{
- return;
+ /* nothing done in this version */
}
static void smpi_check_options(){
xbt_assert(xbt_cfg_get_int("smpi/async-small-thresh") <= xbt_cfg_get_int("smpi/send-is-detached-thresh"));
- if (xbt_cfg_is_default_value("smpi/running-power")) {
+ if (xbt_cfg_is_default_value("smpi/host-speed")) {
XBT_INFO("You did not set the power of the host running the simulation. "
"The timings will certainly not be accurate. "
- "Use the option \"--cfg=smpi/running-power:<flops>\" to set its value."
+ "Use the option \"--cfg=smpi/host-speed:<flops>\" to set its value."
"Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
}
+
+ xbt_assert(xbt_cfg_get_double("smpi/cpu-threshold") >=0,
+ "The 'smpi/cpu-threshold' option cannot have negative values [anymore]. If you want to discard "
+ "the simulation of any computation, please use 'smpi/simulate-computation:no' instead.");
}
int smpi_enabled() {
- return process_data != NULL;
+ return process_data != nullptr;
}
void smpi_global_init()
}
}
+#if HAVE_PAPI
+ // This map holds for each computation unit (such as "default" or "process1" etc.)
+ // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
+ // and the (computed) event_set.
+ std::map</* computation unit name */ std::string, papi_process_data> units2papi_setup;
+
+ if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
+ if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
+ XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?"
+ " Expected version is %i",
+ PAPI_VER_CURRENT);
+
+ typedef boost::tokenizer<boost::char_separator<char>> Tokenizer;
+ boost::char_separator<char> separator_units(";");
+ std::string str = std::string(xbt_cfg_get_string("smpi/papi-events"));
+ Tokenizer tokens(str, separator_units);
+
+ // Iterate over all the computational units. This could be
+ // processes, hosts, threads, ranks... You name it. I'm not exactly
+ // sure what we will support eventually, so I'll leave it at the
+ // general term "units".
+ for (auto& unit_it : tokens) {
+ boost::char_separator<char> separator_events(":");
+ Tokenizer event_tokens(unit_it, separator_events);
+
+ int event_set = PAPI_NULL;
+ if (PAPI_create_eventset(&event_set) != PAPI_OK) {
+ // TODO: Should this let the whole simulation die?
+ XBT_CRITICAL("Could not create PAPI event set during init.");
+ }
+
+ // NOTE: We cannot use a map here, as we must obey the order of the counters
+ // This is important for PAPI: We need to map the values of counters back
+ // to the event_names (so, when PAPI_read() has finished)!
+ papi_counter_t counters2values;
+
+ // Iterate over all counters that were specified for this specific
+ // unit.
+ // Note that we need to remove the name of the unit
+ // (that could also be the "default" value), which always comes first.
+ // Hence, we start at ++(events.begin())!
+ for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); events_it++) {
+
+ int event_code = PAPI_NULL;
+ char* event_name = const_cast<char*>((*events_it).c_str());
+ if (PAPI_event_name_to_code(event_name, &event_code) == PAPI_OK) {
+ if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
+ XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
+ continue;
+ } else {
+ XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
+ }
+ } else {
+ XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
+ continue;
+ }
+
+ counters2values.push_back(
+ // We cannot just pass *events_it, as this is of type const basic_string
+ std::make_pair<std::string, long long>(std::string(*events_it), 0));
+ }
+
+ std::string unit_name = *(event_tokens.begin());
+ papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
+
+ units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
+ }
+ }
+#endif
if (process_count == 0){
process_count = SIMIX_process_count();
smpirun=1;
}
smpi_universe_size = process_count;
- process_data = xbt_new0(smpi_process_data_t, process_count);
+ process_data = new smpi_process_data_t[process_count];
for (i = 0; i < process_count; i++) {
- process_data[i] = xbt_new(s_smpi_process_data_t, 1);
- process_data[i]->argc = NULL;
- process_data[i]->argv = NULL;
+ process_data[i] = new s_smpi_process_data_t;
+ process_data[i]->argc = nullptr;
+ process_data[i]->argv = nullptr;
process_data[i]->mailbox = simcall_mbox_create(get_mailbox_name(name, i));
process_data[i]->mailbox_small = simcall_mbox_create(get_mailbox_name_small(name, i));
process_data[i]->mailboxes_mutex = xbt_mutex_init();
MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
process_data[i]->comm_self = MPI_COMM_NULL;
process_data[i]->comm_intra = MPI_COMM_NULL;
- process_data[i]->comm_world = NULL;
+ process_data[i]->comm_world = nullptr;
process_data[i]->state = SMPI_UNINITIALIZED;
process_data[i]->sampling = 0;
- process_data[i]->finalization_barrier = NULL;
+ process_data[i]->finalization_barrier = nullptr;
process_data[i]->return_value = 0;
- if (xbt_cfg_get_boolean("smpi/trace-call-location")) {
- process_data[i]->trace_call_loc = xbt_new(smpi_trace_call_location_t, 1);
+#if HAVE_PAPI
+ if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
+ // TODO: Implement host/process/thread based counters. This implementation
+ // just always takes the values passed via "default", like this:
+ // "default:COUNTER1:COUNTER2:COUNTER3;".
+ auto it = units2papi_setup.find(papi_default_config_name);
+ if (it != units2papi_setup.end()) {
+ process_data[i]->papi_event_set = it->second.event_set;
+ process_data[i]->papi_counter_data = it->second.counter_data;
+ XBT_DEBUG("Setting PAPI set for process %i", i);
+ } else {
+ process_data[i]->papi_event_set = PAPI_NULL;
+ XBT_DEBUG("No PAPI set for process %i", i);
+ }
}
+#endif
}
//if the process was launched through smpirun script we generate a global mpi_comm_world
//if not, we let MPI_COMM_NULL, and the comm world will be private to each mpi instance
if(smpirun){
group = smpi_group_new(process_count);
- MPI_COMM_WORLD = smpi_comm_new(group, NULL);
+ MPI_COMM_WORLD = smpi_comm_new(group, nullptr);
MPI_Attr_put(MPI_COMM_WORLD, MPI_UNIVERSE_SIZE, reinterpret_cast<void *>(process_count));
- xbt_bar_t bar=xbt_barrier_init(process_count);
+ msg_bar_t bar = MSG_barrier_init(process_count);
for (i = 0; i < process_count; i++) {
smpi_group_set_mapping(group, i, i);
void smpi_global_destroy()
{
int count = smpi_process_count();
- int i;
smpi_bench_destroy();
if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
while (smpi_group_unuse(smpi_comm_group(MPI_COMM_WORLD)) > 0);
- xbt_barrier_destroy(process_data[0]->finalization_barrier);
+ MSG_barrier_destroy(process_data[0]->finalization_barrier);
}else{
smpi_deployment_cleanup_instances();
}
- for (i = 0; i < count; i++) {
+ for (int i = 0; i < count; i++) {
if(process_data[i]->comm_self!=MPI_COMM_NULL){
smpi_comm_destroy(process_data[i]->comm_self);
}
}
xbt_os_timer_free(process_data[i]->timer);
xbt_mutex_destroy(process_data[i]->mailboxes_mutex);
- if (xbt_cfg_get_boolean("smpi/trace-call-location")) {
- xbt_free(process_data[i]->trace_call_loc);
- }
- xbt_free(process_data[i]);
+ delete process_data[i];
}
- xbt_free(process_data);
- process_data = NULL;
+ delete[] process_data;
+ process_data = nullptr;
if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
smpi_comm_cleanup_smp(MPI_COMM_WORLD);
smpi_comm_cleanup_attributes(MPI_COMM_WORLD);
- if(smpi_coll_cleanup_callback!=NULL)
+ if(smpi_coll_cleanup_callback!=nullptr)
smpi_coll_cleanup_callback();
xbt_free(MPI_COMM_WORLD);
}
}
xbt_free(index_to_process_data);
+ if(smpi_type_keyvals!=nullptr)
+ xbt_dict_free(&smpi_type_keyvals);
+ if(smpi_comm_keyvals!=nullptr)
+ xbt_dict_free(&smpi_comm_keyvals);
if(smpi_privatize_global_variables)
smpi_destroy_global_memory_segments();
smpi_free_static();
void __attribute__ ((weak)) user_main_()
{
xbt_die("Should not be in this smpi_simulated_main");
- return;
}
int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
mpi_coll_barrier_fun = reinterpret_cast<int (*)(MPI_Comm comm)>
(mpi_coll_barrier_description[barrier_id].coll);
- smpi_coll_cleanup_callback=NULL;
+ smpi_coll_cleanup_callback=nullptr;
smpi_cpu_threshold = xbt_cfg_get_double("smpi/cpu-threshold");
- smpi_running_power = xbt_cfg_get_double("smpi/running-power");
+ smpi_host_speed = xbt_cfg_get_double("smpi/host-speed");
smpi_privatize_global_variables = xbt_cfg_get_boolean("smpi/privatize-global-variables");
if (smpi_cpu_threshold < 0)
smpi_cpu_threshold = DBL_MAX;
{
srand(SMPI_RAND_SEED);
- if (getenv("SMPI_PRETEND_CC") != NULL) {
+ if (getenv("SMPI_PRETEND_CC") != nullptr) {
/* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
* configuration tools */
return 0;