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. */
13 #include "private.hpp"
14 #include "simgrid/s4u/Mailbox.hpp"
15 #include "smpi/smpi_shared_malloc.hpp"
16 #include "simgrid/sg_config.h"
17 #include "src/kernel/activity/SynchroComm.hpp"
18 #include "src/mc/mc_record.h"
19 #include "src/mc/mc_replay.h"
20 #include "src/msg/msg_private.h"
21 #include "src/simix/smx_private.h"
22 #include "surf/surf.h"
23 #include "xbt/replay.hpp"
24 #include <xbt/config.hpp>
26 #include <float.h> /* DBL_MAX */
37 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)");
38 #include <boost/tokenizer.hpp>
39 #include <boost/algorithm/string.hpp> /* trim_right / trim_left */
42 /* RTLD_DEEPBIND is a bad idea of GNU ld that obviously does not exist on other platforms
43 * See https://www.akkadia.org/drepper/dsohowto.pdf
44 * and https://lists.freebsd.org/pipermail/freebsd-current/2016-March/060284.html
46 #define RTLD_DEEPBIND 0
49 /* Mac OSX does not have any header file providing that definition so we have to duplicate it here. Bummers. */
50 extern char** environ; /* we use it in posix_spawnp below */
54 const char* papi_default_config_name = "default";
56 struct papi_process_data {
57 papi_counter_t counter_data;
62 std::unordered_map<std::string, double> location2speedup;
64 simgrid::smpi::Process **process_data = nullptr;
65 int process_count = 0;
66 int smpi_universe_size = 0;
67 int* index_to_process_data = nullptr;
68 extern double smpi_total_benched_time;
69 xbt_os_timer_t global_timer;
70 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
71 MPI_Errhandler *MPI_ERRORS_RETURN = nullptr;
72 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = nullptr;
73 MPI_Errhandler *MPI_ERRHANDLER_NULL = nullptr;
74 static simgrid::config::Flag<double> smpi_wtime_sleep(
75 "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0);
76 static simgrid::config::Flag<double> smpi_init_sleep(
77 "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
79 void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t) = &smpi_comm_copy_buffer_callback;
83 int smpi_process_count()
88 simgrid::smpi::Process* smpi_process()
90 simgrid::MsgActorExt* msgExt = static_cast<simgrid::MsgActorExt*>(SIMIX_process_self()->data);
91 return static_cast<simgrid::smpi::Process*>(msgExt->data);
94 simgrid::smpi::Process* smpi_process_remote(int index)
96 return process_data[index_to_process_data[index]];
99 MPI_Comm smpi_process_comm_self(){
100 return smpi_process()->comm_self();
103 void smpi_process_init(int *argc, char ***argv){
104 simgrid::smpi::Process::init(argc, argv);
107 int smpi_process_index(){
108 return smpi_process()->index();
112 int smpi_global_size()
114 char *value = getenv("SMPI_GLOBAL_SIZE");
115 xbt_assert(value,"Please set env var SMPI_GLOBAL_SIZE to the expected number of processes.");
117 return xbt_str_parse_int(value, "SMPI_GLOBAL_SIZE contains a non-numerical value: %s");
120 void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
122 smpi_comm_copy_data_callback = callback;
125 static void print(std::vector<std::pair<size_t, size_t>> vec) {
126 fprintf(stderr, "{");
128 fprintf(stderr, "(0x%lx, 0x%lx),", elt.first, elt.second);
130 fprintf(stderr, "}\n");
132 static void memcpy_private(void *dest, const void *src, size_t n, std::vector<std::pair<size_t, size_t>> &private_blocks) {
133 for(auto block : private_blocks) {
134 memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first);
138 static void check_blocks(std::vector<std::pair<size_t, size_t>> &private_blocks, size_t buff_size) {
139 for(auto block : private_blocks) {
140 xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
144 void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size)
146 simgrid::kernel::activity::Comm *comm = dynamic_cast<simgrid::kernel::activity::Comm*>(synchro);
147 int src_shared=0, dst_shared=0;
148 size_t src_offset=0, dst_offset=0;
149 std::vector<std::pair<size_t, size_t>> src_private_blocks;
150 std::vector<std::pair<size_t, size_t>> dst_private_blocks;
151 XBT_DEBUG("Copy the data over");
152 if((src_shared=smpi_is_shared(buff, src_private_blocks, &src_offset))) {
153 XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff);
154 src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
157 src_private_blocks.clear();
158 src_private_blocks.push_back(std::make_pair(0, buff_size));
160 if((dst_shared=smpi_is_shared((char*)comm->dst_buff, dst_private_blocks, &dst_offset))) {
161 XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff);
162 dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
165 dst_private_blocks.clear();
166 dst_private_blocks.push_back(std::make_pair(0, buff_size));
168 check_blocks(src_private_blocks, buff_size);
169 check_blocks(dst_private_blocks, buff_size);
170 auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
171 check_blocks(private_blocks, buff_size);
173 if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && (static_cast<char*>(buff) >= smpi_start_data_exe)
174 && (static_cast<char*>(buff) < smpi_start_data_exe + smpi_size_data_exe )
176 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
178 smpi_switch_data_segment(
179 (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->src_proc->data)->data))->index()));
180 tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
181 memcpy_private(tmpbuff, buff, buff_size, private_blocks);
184 if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_start_data_exe)
185 && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){
186 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
187 smpi_switch_data_segment(
188 (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->dst_proc->data)->data))->index()));
190 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
191 memcpy_private(comm->dst_buff, tmpbuff, buff_size, private_blocks);
193 if (comm->detached) {
194 // if this is a detached send, the source buffer was duplicated by SMPI
195 // sender to make the original buffer available to the application ASAP
197 //It seems that the request is used after the call there this should be free somewhere else but where???
198 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
199 comm->src_buff = nullptr;
201 if(tmpbuff!=buff)xbt_free(tmpbuff);
205 void smpi_comm_null_copy_buffer_callback(smx_activity_t comm, void *buff, size_t buff_size)
207 /* nothing done in this version */
210 static void smpi_check_options(){
211 //check correctness of MPI parameters
213 xbt_assert(xbt_cfg_get_int("smpi/async-small-thresh") <= xbt_cfg_get_int("smpi/send-is-detached-thresh"));
215 if (xbt_cfg_is_default_value("smpi/host-speed")) {
216 XBT_INFO("You did not set the power of the host running the simulation. "
217 "The timings will certainly not be accurate. "
218 "Use the option \"--cfg=smpi/host-speed:<flops>\" to set its value."
219 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
222 xbt_assert(xbt_cfg_get_double("smpi/cpu-threshold") >=0,
223 "The 'smpi/cpu-threshold' option cannot have negative values [anymore]. If you want to discard "
224 "the simulation of any computation, please use 'smpi/simulate-computation:no' instead.");
228 return process_data != nullptr;
231 void smpi_global_init()
235 if (!MC_is_active()) {
236 global_timer = xbt_os_timer_new();
237 xbt_os_walltimer_start(global_timer);
240 if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') {
241 std::string filename {xbt_cfg_get_string("smpi/comp-adjustment-file")};
242 std::ifstream fstream(filename);
243 if (!fstream.is_open()) {
244 xbt_die("Could not open file %s. Does it exist?", filename.c_str());
248 typedef boost::tokenizer< boost::escaped_list_separator<char>> Tokenizer;
249 std::getline(fstream, line); // Skip the header line
250 while (std::getline(fstream, line)) {
252 Tokenizer::iterator it = tok.begin();
253 Tokenizer::iterator end = std::next(tok.begin());
255 std::string location = *it;
256 boost::trim(location);
257 location2speedup.insert(std::pair<std::string, double>(location, std::stod(*end)));
262 // This map holds for each computation unit (such as "default" or "process1" etc.)
263 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
264 // and the (computed) event_set.
265 std::map</* computation unit name */ std::string, papi_process_data> units2papi_setup;
267 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
268 if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
269 XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?"
270 " Expected version is %i",
273 typedef boost::tokenizer<boost::char_separator<char>> Tokenizer;
274 boost::char_separator<char> separator_units(";");
275 std::string str = std::string(xbt_cfg_get_string("smpi/papi-events"));
276 Tokenizer tokens(str, separator_units);
278 // Iterate over all the computational units. This could be
279 // processes, hosts, threads, ranks... You name it. I'm not exactly
280 // sure what we will support eventually, so I'll leave it at the
281 // general term "units".
282 for (auto& unit_it : tokens) {
283 boost::char_separator<char> separator_events(":");
284 Tokenizer event_tokens(unit_it, separator_events);
286 int event_set = PAPI_NULL;
287 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
288 // TODO: Should this let the whole simulation die?
289 XBT_CRITICAL("Could not create PAPI event set during init.");
292 // NOTE: We cannot use a map here, as we must obey the order of the counters
293 // This is important for PAPI: We need to map the values of counters back
294 // to the event_names (so, when PAPI_read() has finished)!
295 papi_counter_t counters2values;
297 // Iterate over all counters that were specified for this specific
299 // Note that we need to remove the name of the unit
300 // (that could also be the "default" value), which always comes first.
301 // Hence, we start at ++(events.begin())!
302 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); events_it++) {
304 int event_code = PAPI_NULL;
305 char* event_name = const_cast<char*>((*events_it).c_str());
306 if (PAPI_event_name_to_code(event_name, &event_code) == PAPI_OK) {
307 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
308 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
311 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
314 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
318 counters2values.push_back(
319 // We cannot just pass *events_it, as this is of type const basic_string
320 std::make_pair<std::string, long long>(std::string(*events_it), 0));
323 std::string unit_name = *(event_tokens.begin());
324 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
326 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
332 msg_bar_t finalization_barrier = nullptr;
333 if (process_count == 0){
334 process_count = SIMIX_process_count();
336 finalization_barrier = MSG_barrier_init(process_count);
338 smpi_universe_size = process_count;
339 process_data = new simgrid::smpi::Process*[process_count];
340 for (int i = 0; i < process_count; i++) {
341 process_data[i] = new simgrid::smpi::Process(i, finalization_barrier);
343 //if the process was launched through smpirun script we generate a global mpi_comm_world
344 //if not, we let MPI_COMM_NULL, and the comm world will be private to each mpi instance
346 group = new simgrid::smpi::Group(process_count);
347 MPI_COMM_WORLD = new simgrid::smpi::Comm(group, nullptr);
348 MPI_Attr_put(MPI_COMM_WORLD, MPI_UNIVERSE_SIZE, reinterpret_cast<void *>(process_count));
350 for (int i = 0; i < process_count; i++)
351 group->set_mapping(i, i);
355 void smpi_global_destroy()
357 int count = smpi_process_count();
359 smpi_bench_destroy();
360 smpi_shared_destroy();
361 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
362 delete MPI_COMM_WORLD->group();
363 MSG_barrier_destroy(process_data[0]->finalization_barrier());
365 smpi_deployment_cleanup_instances();
367 for (int i = 0; i < count; i++) {
368 if(process_data[i]->comm_self()!=MPI_COMM_NULL){
369 simgrid::smpi::Comm::destroy(process_data[i]->comm_self());
371 if(process_data[i]->comm_intra()!=MPI_COMM_NULL){
372 simgrid::smpi::Comm::destroy(process_data[i]->comm_intra());
374 xbt_os_timer_free(process_data[i]->timer());
375 xbt_mutex_destroy(process_data[i]->mailboxes_mutex());
376 delete process_data[i];
378 delete[] process_data;
379 process_data = nullptr;
381 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
382 MPI_COMM_WORLD->cleanup_smp();
383 MPI_COMM_WORLD->cleanup_attr<simgrid::smpi::Comm>();
384 if(simgrid::smpi::Colls::smpi_coll_cleanup_callback!=nullptr)
385 simgrid::smpi::Colls::smpi_coll_cleanup_callback();
386 delete MPI_COMM_WORLD;
389 MPI_COMM_WORLD = MPI_COMM_NULL;
391 if (!MC_is_active()) {
392 xbt_os_timer_free(global_timer);
395 xbt_free(index_to_process_data);
396 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
397 smpi_destroy_global_memory_segments();
403 static void smpi_init_logs(){
405 /* Connect log categories. See xbt/log.c */
407 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this function: xbt_log_appender_file.c depends on it
408 DO NOT connect this in XBT or so, or it will be useless to xbt_log_appender_file.c */
409 XBT_LOG_CONNECT(instr_smpi);
410 XBT_LOG_CONNECT(smpi_bench);
411 XBT_LOG_CONNECT(smpi_coll);
412 XBT_LOG_CONNECT(smpi_colls);
413 XBT_LOG_CONNECT(smpi_comm);
414 XBT_LOG_CONNECT(smpi_datatype);
415 XBT_LOG_CONNECT(smpi_dvfs);
416 XBT_LOG_CONNECT(smpi_group);
417 XBT_LOG_CONNECT(smpi_kernel);
418 XBT_LOG_CONNECT(smpi_mpi);
419 XBT_LOG_CONNECT(smpi_memory);
420 XBT_LOG_CONNECT(smpi_op);
421 XBT_LOG_CONNECT(smpi_pmpi);
422 XBT_LOG_CONNECT(smpi_request);
423 XBT_LOG_CONNECT(smpi_replay);
424 XBT_LOG_CONNECT(smpi_rma);
425 XBT_LOG_CONNECT(smpi_shared);
426 XBT_LOG_CONNECT(smpi_utils);
430 static void smpi_init_options(){
432 simgrid::smpi::Colls::set_collectives();
433 simgrid::smpi::Colls::smpi_coll_cleanup_callback=nullptr;
434 smpi_cpu_threshold = xbt_cfg_get_double("smpi/cpu-threshold");
435 smpi_host_speed = xbt_cfg_get_double("smpi/host-speed");
436 const char* smpi_privatize_option = xbt_cfg_get_string("smpi/privatize-global-variables");
437 if (std::strcmp(smpi_privatize_option, "no") == 0)
438 smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE;
439 else if (std::strcmp(smpi_privatize_option, "yes") == 0)
440 smpi_privatize_global_variables = SMPI_PRIVATIZE_DEFAULT;
441 else if (std::strcmp(smpi_privatize_option, "mmap") == 0)
442 smpi_privatize_global_variables = SMPI_PRIVATIZE_MMAP;
443 else if (std::strcmp(smpi_privatize_option, "dlopen") == 0)
444 smpi_privatize_global_variables = SMPI_PRIVATIZE_DLOPEN;
446 // Some compatibility stuff:
447 else if (std::strcmp(smpi_privatize_option, "1") == 0)
448 smpi_privatize_global_variables = SMPI_PRIVATIZE_DEFAULT;
449 else if (std::strcmp(smpi_privatize_option, "0") == 0)
450 smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE;
453 xbt_die("Invalid value for smpi/privatize-global-variables: %s",
454 smpi_privatize_option);
456 if (smpi_cpu_threshold < 0)
457 smpi_cpu_threshold = DBL_MAX;
459 char* val = xbt_cfg_get_string("smpi/shared-malloc");
460 if (!strcasecmp(val, "yes") || !strcmp(val, "1") || !strcasecmp(val, "on") || !strcasecmp(val, "global")) {
461 smpi_cfg_shared_malloc = shmalloc_global;
462 } else if (!strcasecmp(val, "local")) {
463 smpi_cfg_shared_malloc = shmalloc_local;
464 } else if (!strcasecmp(val, "no") || !strcmp(val, "0") || !strcasecmp(val, "off")) {
465 smpi_cfg_shared_malloc = shmalloc_none;
467 xbt_die("Invalid value '%s' for option smpi/shared-malloc. Possible values: 'on' or 'global', 'local', 'off'",
472 static int execute_command(const char * const argv[])
476 if (posix_spawnp(&pid, argv[0], nullptr, nullptr, (char* const*) argv, environ) != 0)
478 if (waitpid(pid, &status, 0) != pid)
483 typedef std::function<int(int argc, char *argv[])> smpi_entry_point_type;
484 typedef int (* smpi_c_entry_point_type)(int argc, char **argv);
485 typedef void (* smpi_fortran_entry_point_type)(void);
487 static int smpi_run_entry_point(smpi_entry_point_type entry_point, std::vector<std::string> args)
489 const int argc = args.size();
490 std::unique_ptr<char*[]> argv(new char*[argc + 1]);
491 for (int i = 0; i != argc; ++i)
492 argv[i] = args[i].empty() ? const_cast<char*>(""): &args[i].front();
493 argv[argc] = nullptr;
495 int res = entry_point(argc, argv.get());
497 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
498 smpi_process()->set_return_value(res);
503 // TODO, remove the number of functions involved here
504 static smpi_entry_point_type smpi_resolve_function(void* handle)
506 smpi_fortran_entry_point_type entry_point2 =
507 (smpi_fortran_entry_point_type) dlsym(handle, "user_main_");
508 if (entry_point2 != nullptr) {
509 // fprintf(stderr, "EP user_main_=%p\n", entry_point2);
510 return [entry_point2](int argc, char** argv) {
511 smpi_process_init(&argc, &argv);
517 smpi_c_entry_point_type entry_point = (smpi_c_entry_point_type) dlsym(handle, "main");
518 if (entry_point != nullptr) {
519 // fprintf(stderr, "EP main=%p\n", entry_point);
523 return smpi_entry_point_type();
526 int smpi_main(const char* executable, int argc, char *argv[])
528 srand(SMPI_RAND_SEED);
530 if (getenv("SMPI_PRETEND_CC") != nullptr) {
531 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
532 * configuration tools */
537 TRACE_global_init(&argc, argv);
538 TRACE_add_start_function(TRACE_smpi_alloc);
539 TRACE_add_end_function(TRACE_smpi_release);
541 SIMIX_global_init(&argc, argv);
542 MSG_init(&argc,argv);
544 SMPI_switch_data_segment = &smpi_switch_data_segment;
548 // parse the platform file: get the host list
549 SIMIX_create_environment(argv[1]);
550 SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback);
552 static std::size_t rank = 0;
554 if (smpi_privatize_global_variables == SMPI_PRIVATIZE_DLOPEN) {
556 std::string executable_copy = executable;
557 simix_global->default_function = [executable_copy](std::vector<std::string> args) {
558 return std::function<void()>([executable_copy, args] {
560 // Copy the dynamic library:
561 std::string target_executable = executable_copy
562 + "_" + std::to_string(getpid())
563 + "_" + std::to_string(rank++) + ".so";
564 // TODO, execute directly instead of relying on cp
565 const char* command1 [] = {
566 "cp", "--reflink=auto", "--", executable_copy.c_str(), target_executable.c_str(),
569 const char* command2 [] = {
570 "cp", "--", executable_copy.c_str(), target_executable.c_str(),
573 if (execute_command(command1) != 0 && execute_command(command2) != 0)
574 xbt_die("copy failed");
576 // Load the copy and resolve the entry point:
577 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);
578 unlink(target_executable.c_str());
579 if (handle == nullptr)
580 xbt_die("dlopen failed");
581 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
583 xbt_die("Could not resolve entry point");
585 smpi_run_entry_point(entry_point, args);
592 // Load the dynamic library and resolve the entry point:
593 void* handle = dlopen(executable, RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);
594 if (handle == nullptr)
595 xbt_die("dlopen failed for %s", executable);
596 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
598 xbt_die("main not found in %s", executable);
599 // TODO, register the executable for SMPI privatization
601 // Execute the same entry point for each simulated process:
602 simix_global->default_function = [entry_point](std::vector<std::string> args) {
603 return std::function<void()>([entry_point, args] {
604 smpi_run_entry_point(entry_point, args);
610 SIMIX_launch_application(argv[2]);
614 smpi_check_options();
616 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
617 smpi_initialize_global_memory_segments();
619 /* Clean IO before the run */
623 if (MC_is_active()) {
629 xbt_os_walltimer_stop(global_timer);
630 if (xbt_cfg_get_boolean("smpi/display-timing")){
631 double global_time = xbt_os_timer_elapsed(global_timer);
632 XBT_INFO("Simulated time: %g seconds. \n\n"
633 "The simulation took %g seconds (after parsing and platform setup)\n"
634 "%g seconds were actual computation of the application",
635 SIMIX_get_clock(), global_time , smpi_total_benched_time);
637 if (smpi_total_benched_time/global_time>=0.75)
638 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
639 "You may want to use sampling functions or trace replay to reduce this.");
642 int count = smpi_process_count();
644 for (i = 0; i < count; i++) {
645 if(process_data[i]->return_value()!=0){
646 ret=process_data[i]->return_value();//return first non 0 value
650 smpi_global_destroy();
657 // This function can be called from extern file, to initialize logs, options, and processes of smpi
658 // without the need of smpirun
663 smpi_check_options();
664 if (TRACE_is_enabled() && TRACE_is_configured())
666 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
667 smpi_initialize_global_memory_segments();
670 void SMPI_finalize(){
671 smpi_global_destroy();
674 void smpi_mpi_init() {
675 if(smpi_init_sleep > 0)
676 simcall_process_sleep(smpi_init_sleep);
679 double smpi_mpi_wtime(){
681 if (smpi_process()->initialized() != 0 && smpi_process()->finalized() == 0 && smpi_process()->sampling() == 0) {
683 time = SIMIX_get_clock();
684 // to avoid deadlocks if used as a break condition, such as
685 // while (MPI_Wtime(...) < time_limit) {
688 // because the time will not normally advance when only calls to MPI_Wtime
689 // are made -> deadlock (MPI_Wtime never reaches the time limit)
690 if(smpi_wtime_sleep > 0)
691 simcall_process_sleep(smpi_wtime_sleep);
694 time = SIMIX_get_clock();