1 /* Copyright (c) 2007-2017. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
9 #include "simgrid/s4u/Mailbox.hpp"
10 #include "simgrid/sg_config.h"
11 #include "src/kernel/activity/SynchroComm.hpp"
12 #include "src/mc/mc_record.h"
13 #include "src/mc/mc_replay.h"
14 #include "src/msg/msg_private.h"
15 #include "src/simix/smx_private.h"
16 #include "surf/surf.h"
17 #include "xbt/replay.hpp"
18 #include <xbt/config.hpp>
20 #include <float.h> /* DBL_MAX */
29 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)");
30 #include <boost/tokenizer.hpp>
31 #include <boost/algorithm/string.hpp> /* trim_right / trim_left */
35 const char* papi_default_config_name = "default";
37 struct papi_process_data {
38 papi_counter_t counter_data;
43 std::unordered_map<std::string, double> location2speedup;
45 simgrid::smpi::Process **process_data = nullptr;
46 int process_count = 0;
47 int smpi_universe_size = 0;
48 int* index_to_process_data = nullptr;
49 extern double smpi_total_benched_time;
50 xbt_os_timer_t global_timer;
51 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
52 MPI_Errhandler *MPI_ERRORS_RETURN = nullptr;
53 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = nullptr;
54 MPI_Errhandler *MPI_ERRHANDLER_NULL = nullptr;
55 static simgrid::config::Flag<double> smpi_wtime_sleep(
56 "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0);
57 static simgrid::config::Flag<double> smpi_init_sleep(
58 "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
60 void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t) = &smpi_comm_copy_buffer_callback;
64 int smpi_process_count()
69 simgrid::smpi::Process* smpi_process()
71 simgrid::MsgActorExt* msgExt = static_cast<simgrid::MsgActorExt*>(SIMIX_process_self()->data);
72 return static_cast<simgrid::smpi::Process*>(msgExt->data);
75 simgrid::smpi::Process* smpi_process_remote(int index)
77 return process_data[index_to_process_data[index]];
80 MPI_Comm smpi_process_comm_self(){
81 return smpi_process()->comm_self();
84 void smpi_process_init(int *argc, char ***argv){
85 simgrid::smpi::Process::init(argc, argv);
88 int smpi_process_index(){
89 return smpi_process()->index();
93 int smpi_global_size()
95 char *value = getenv("SMPI_GLOBAL_SIZE");
96 xbt_assert(value,"Please set env var SMPI_GLOBAL_SIZE to the expected number of processes.");
98 return xbt_str_parse_int(value, "SMPI_GLOBAL_SIZE contains a non-numerical value: %s");
101 void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
103 smpi_comm_copy_data_callback = callback;
106 void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size)
109 simgrid::kernel::activity::Comm *comm = dynamic_cast<simgrid::kernel::activity::Comm*>(synchro);
111 XBT_DEBUG("Copy the data over");
112 if(smpi_is_shared(buff)){
113 XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff);
114 }else if(smpi_is_shared((char*)comm->dst_buff)){
115 XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff);
118 if((smpi_privatize_global_variables) && (static_cast<char*>(buff) >= smpi_start_data_exe)
119 && (static_cast<char*>(buff) < smpi_start_data_exe + smpi_size_data_exe )
121 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
123 smpi_switch_data_segment(
124 (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->src_proc->data)->data))->index()));
125 tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
126 memcpy(tmpbuff, buff, buff_size);
129 if((smpi_privatize_global_variables) && ((char*)comm->dst_buff >= smpi_start_data_exe)
130 && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){
131 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
132 smpi_switch_data_segment(
133 (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->dst_proc->data)->data))->index()));
136 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
137 memcpy(comm->dst_buff, tmpbuff, buff_size);
139 if (comm->detached) {
140 // if this is a detached send, the source buffer was duplicated by SMPI
141 // sender to make the original buffer available to the application ASAP
143 //It seems that the request is used after the call there this should be free somewhere else but where???
144 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
145 comm->src_buff = nullptr;
147 if(tmpbuff!=buff)xbt_free(tmpbuff);
152 void smpi_comm_null_copy_buffer_callback(smx_activity_t comm, void *buff, size_t buff_size)
154 /* nothing done in this version */
157 static void smpi_check_options(){
158 //check correctness of MPI parameters
160 xbt_assert(xbt_cfg_get_int("smpi/async-small-thresh") <= xbt_cfg_get_int("smpi/send-is-detached-thresh"));
162 if (xbt_cfg_is_default_value("smpi/host-speed")) {
163 XBT_INFO("You did not set the power of the host running the simulation. "
164 "The timings will certainly not be accurate. "
165 "Use the option \"--cfg=smpi/host-speed:<flops>\" to set its value."
166 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
169 xbt_assert(xbt_cfg_get_double("smpi/cpu-threshold") >=0,
170 "The 'smpi/cpu-threshold' option cannot have negative values [anymore]. If you want to discard "
171 "the simulation of any computation, please use 'smpi/simulate-computation:no' instead.");
175 return process_data != nullptr;
178 void smpi_global_init()
184 if (!MC_is_active()) {
185 global_timer = xbt_os_timer_new();
186 xbt_os_walltimer_start(global_timer);
189 if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') {
190 std::string filename {xbt_cfg_get_string("smpi/comp-adjustment-file")};
191 std::ifstream fstream(filename);
192 if (!fstream.is_open()) {
193 xbt_die("Could not open file %s. Does it exist?", filename.c_str());
197 typedef boost::tokenizer< boost::escaped_list_separator<char>> Tokenizer;
198 std::getline(fstream, line); // Skip the header line
199 while (std::getline(fstream, line)) {
201 Tokenizer::iterator it = tok.begin();
202 Tokenizer::iterator end = std::next(tok.begin());
204 std::string location = *it;
205 boost::trim(location);
206 location2speedup.insert(std::pair<std::string, double>(location, std::stod(*end)));
211 // This map holds for each computation unit (such as "default" or "process1" etc.)
212 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
213 // and the (computed) event_set.
214 std::map</* computation unit name */ std::string, papi_process_data> units2papi_setup;
216 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
217 if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
218 XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?"
219 " Expected version is %i",
222 typedef boost::tokenizer<boost::char_separator<char>> Tokenizer;
223 boost::char_separator<char> separator_units(";");
224 std::string str = std::string(xbt_cfg_get_string("smpi/papi-events"));
225 Tokenizer tokens(str, separator_units);
227 // Iterate over all the computational units. This could be
228 // processes, hosts, threads, ranks... You name it. I'm not exactly
229 // sure what we will support eventually, so I'll leave it at the
230 // general term "units".
231 for (auto& unit_it : tokens) {
232 boost::char_separator<char> separator_events(":");
233 Tokenizer event_tokens(unit_it, separator_events);
235 int event_set = PAPI_NULL;
236 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
237 // TODO: Should this let the whole simulation die?
238 XBT_CRITICAL("Could not create PAPI event set during init.");
241 // NOTE: We cannot use a map here, as we must obey the order of the counters
242 // This is important for PAPI: We need to map the values of counters back
243 // to the event_names (so, when PAPI_read() has finished)!
244 papi_counter_t counters2values;
246 // Iterate over all counters that were specified for this specific
248 // Note that we need to remove the name of the unit
249 // (that could also be the "default" value), which always comes first.
250 // Hence, we start at ++(events.begin())!
251 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); events_it++) {
253 int event_code = PAPI_NULL;
254 char* event_name = const_cast<char*>((*events_it).c_str());
255 if (PAPI_event_name_to_code(event_name, &event_code) == PAPI_OK) {
256 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
257 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
260 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
263 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
267 counters2values.push_back(
268 // We cannot just pass *events_it, as this is of type const basic_string
269 std::make_pair<std::string, long long>(std::string(*events_it), 0));
272 std::string unit_name = *(event_tokens.begin());
273 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
275 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
279 if (process_count == 0){
280 process_count = SIMIX_process_count();
283 smpi_universe_size = process_count;
284 process_data = new simgrid::smpi::Process*[process_count];
285 for (i = 0; i < process_count; i++) {
286 process_data[i] = new simgrid::smpi::Process(i);
288 //if the process was launched through smpirun script we generate a global mpi_comm_world
289 //if not, we let MPI_COMM_NULL, and the comm world will be private to each mpi instance
291 group = new simgrid::smpi::Group(process_count);
292 MPI_COMM_WORLD = new simgrid::smpi::Comm(group, nullptr);
293 MPI_Attr_put(MPI_COMM_WORLD, MPI_UNIVERSE_SIZE, reinterpret_cast<void *>(process_count));
294 msg_bar_t bar = MSG_barrier_init(process_count);
296 for (i = 0; i < process_count; i++) {
297 group->set_mapping(i, i);
298 process_data[i]->set_finalization_barrier(bar);
303 void smpi_global_destroy()
305 int count = smpi_process_count();
307 smpi_bench_destroy();
308 smpi_shared_destroy();
309 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
310 delete MPI_COMM_WORLD->group();
311 MSG_barrier_destroy(process_data[0]->finalization_barrier());
313 smpi_deployment_cleanup_instances();
315 for (int i = 0; i < count; i++) {
316 if(process_data[i]->comm_self()!=MPI_COMM_NULL){
317 simgrid::smpi::Comm::destroy(process_data[i]->comm_self());
319 if(process_data[i]->comm_intra()!=MPI_COMM_NULL){
320 simgrid::smpi::Comm::destroy(process_data[i]->comm_intra());
322 xbt_os_timer_free(process_data[i]->timer());
323 xbt_mutex_destroy(process_data[i]->mailboxes_mutex());
324 delete process_data[i];
326 delete[] process_data;
327 process_data = nullptr;
329 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
330 MPI_COMM_WORLD->cleanup_smp();
331 MPI_COMM_WORLD->cleanup_attr<simgrid::smpi::Comm>();
332 if(simgrid::smpi::Colls::smpi_coll_cleanup_callback!=nullptr)
333 simgrid::smpi::Colls::smpi_coll_cleanup_callback();
334 delete MPI_COMM_WORLD;
337 MPI_COMM_WORLD = MPI_COMM_NULL;
339 if (!MC_is_active()) {
340 xbt_os_timer_free(global_timer);
343 xbt_free(index_to_process_data);
344 if(smpi_privatize_global_variables)
345 smpi_destroy_global_memory_segments();
353 void __attribute__ ((weak)) user_main_()
355 xbt_die("Should not be in this smpi_simulated_main");
358 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
360 simgrid::smpi::Process::init(&argc, &argv);
365 inline static int smpi_main_wrapper(int argc, char **argv){
366 int ret = smpi_simulated_main_(argc,argv);
368 XBT_WARN("SMPI process did not return 0. Return value : %d", ret);
369 smpi_process()->set_return_value(ret);
374 int __attribute__ ((weak)) main(int argc, char **argv)
376 return smpi_main(smpi_main_wrapper, argc, argv);
381 static void smpi_init_logs(){
383 /* Connect log categories. See xbt/log.c */
385 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this function: xbt_log_appender_file.c depends on it
386 DO NOT connect this in XBT or so, or it will be useless to xbt_log_appender_file.c */
387 XBT_LOG_CONNECT(instr_smpi);
388 XBT_LOG_CONNECT(smpi_bench);
389 XBT_LOG_CONNECT(smpi_coll);
390 XBT_LOG_CONNECT(smpi_colls);
391 XBT_LOG_CONNECT(smpi_comm);
392 XBT_LOG_CONNECT(smpi_datatype);
393 XBT_LOG_CONNECT(smpi_dvfs);
394 XBT_LOG_CONNECT(smpi_group);
395 XBT_LOG_CONNECT(smpi_kernel);
396 XBT_LOG_CONNECT(smpi_mpi);
397 XBT_LOG_CONNECT(smpi_memory);
398 XBT_LOG_CONNECT(smpi_op);
399 XBT_LOG_CONNECT(smpi_pmpi);
400 XBT_LOG_CONNECT(smpi_request);
401 XBT_LOG_CONNECT(smpi_replay);
402 XBT_LOG_CONNECT(smpi_rma);
403 XBT_LOG_CONNECT(smpi_shared);
404 XBT_LOG_CONNECT(smpi_utils);
408 static void smpi_init_options(){
410 simgrid::smpi::Colls::set_collectives();
411 simgrid::smpi::Colls::smpi_coll_cleanup_callback=nullptr;
412 smpi_cpu_threshold = xbt_cfg_get_double("smpi/cpu-threshold");
413 smpi_host_speed = xbt_cfg_get_double("smpi/host-speed");
414 smpi_privatize_global_variables = xbt_cfg_get_boolean("smpi/privatize-global-variables");
415 if (smpi_cpu_threshold < 0)
416 smpi_cpu_threshold = DBL_MAX;
418 char* val = xbt_cfg_get_string("smpi/shared-malloc");
419 if (!strcasecmp(val, "yes") || !strcmp(val, "1") || !strcasecmp(val, "on") || !strcasecmp(val, "global")) {
420 smpi_cfg_shared_malloc = shmalloc_global;
421 } else if (!strcasecmp(val, "local")) {
422 smpi_cfg_shared_malloc = shmalloc_local;
423 } else if (!strcasecmp(val, "no") || !strcmp(val, "0") || !strcasecmp(val, "off")) {
424 smpi_cfg_shared_malloc = shmalloc_none;
426 xbt_die("Invalid value '%s' for option smpi/shared-malloc. Possible values: 'on' or 'global', 'local', 'off'",
431 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
433 srand(SMPI_RAND_SEED);
435 if (getenv("SMPI_PRETEND_CC") != nullptr) {
436 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
437 * configuration tools */
442 TRACE_global_init(&argc, argv);
443 TRACE_add_start_function(TRACE_smpi_alloc);
444 TRACE_add_end_function(TRACE_smpi_release);
446 SIMIX_global_init(&argc, argv);
447 MSG_init(&argc,argv);
449 SMPI_switch_data_segment = &smpi_switch_data_segment;
453 // parse the platform file: get the host list
454 SIMIX_create_environment(argv[1]);
455 SIMIX_comm_set_copy_data_callback(smpi_comm_copy_data_callback);
456 SIMIX_function_register_default(realmain);
457 SIMIX_launch_application(argv[2]);
461 smpi_check_options();
463 if(smpi_privatize_global_variables)
464 smpi_initialize_global_memory_segments();
466 /* Clean IO before the run */
470 if (MC_is_active()) {
476 xbt_os_walltimer_stop(global_timer);
477 if (xbt_cfg_get_boolean("smpi/display-timing")){
478 double global_time = xbt_os_timer_elapsed(global_timer);
479 XBT_INFO("Simulated time: %g seconds. \n\n"
480 "The simulation took %g seconds (after parsing and platform setup)\n"
481 "%g seconds were actual computation of the application",
482 SIMIX_get_clock(), global_time , smpi_total_benched_time);
484 if (smpi_total_benched_time/global_time>=0.75)
485 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
486 "You may want to use sampling functions or trace replay to reduce this.");
489 int count = smpi_process_count();
491 for (i = 0; i < count; i++) {
492 if(process_data[i]->return_value()!=0){
493 ret=process_data[i]->return_value();//return first non 0 value
497 smpi_global_destroy();
504 // This function can be called from extern file, to initialize logs, options, and processes of smpi
505 // without the need of smpirun
510 smpi_check_options();
511 if (TRACE_is_enabled() && TRACE_is_configured())
513 if(smpi_privatize_global_variables)
514 smpi_initialize_global_memory_segments();
517 void SMPI_finalize(){
518 smpi_global_destroy();
521 void smpi_mpi_init() {
522 if(smpi_init_sleep > 0)
523 simcall_process_sleep(smpi_init_sleep);
526 double smpi_mpi_wtime(){
528 if (smpi_process()->initialized() != 0 && smpi_process()->finalized() == 0 && smpi_process()->sampling() == 0) {
530 time = SIMIX_get_clock();
531 // to avoid deadlocks if used as a break condition, such as
532 // while (MPI_Wtime(...) < time_limit) {
535 // because the time will not normally advance when only calls to MPI_Wtime
536 // are made -> deadlock (MPI_Wtime never reaches the time limit)
537 if(smpi_wtime_sleep > 0)
538 simcall_process_sleep(smpi_wtime_sleep);
541 time = SIMIX_get_clock();