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 "simgrid/sg_config.h"
16 #include "src/kernel/activity/SynchroComm.hpp"
17 #include "src/mc/mc_record.h"
18 #include "src/mc/mc_replay.h"
19 #include "src/msg/msg_private.h"
20 #include "src/simix/smx_private.h"
21 #include "surf/surf.h"
22 #include "xbt/replay.hpp"
23 #include <xbt/config.hpp>
25 #include <float.h> /* DBL_MAX */
36 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)");
37 #include <boost/tokenizer.hpp>
38 #include <boost/algorithm/string.hpp> /* trim_right / trim_left */
41 /* RTLD_DEEPBIND is a bad idea of GNU ld that obviously does not exist on other platforms
42 * See https://www.akkadia.org/drepper/dsohowto.pdf
43 * and https://lists.freebsd.org/pipermail/freebsd-current/2016-March/060284.html
45 #define RTLD_DEEPBIND 0
48 /* Mac OSX does not have any header file providing that definition so we have to duplicate it here. Bummers. */
49 extern char** environ; /* we use it in posix_spawnp below */
53 const char* papi_default_config_name = "default";
55 struct papi_process_data {
56 papi_counter_t counter_data;
61 std::unordered_map<std::string, double> location2speedup;
63 simgrid::smpi::Process **process_data = nullptr;
64 int process_count = 0;
65 int smpi_universe_size = 0;
66 int* index_to_process_data = nullptr;
67 extern double smpi_total_benched_time;
68 xbt_os_timer_t global_timer;
69 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
70 MPI_Errhandler *MPI_ERRORS_RETURN = nullptr;
71 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = nullptr;
72 MPI_Errhandler *MPI_ERRHANDLER_NULL = nullptr;
73 static simgrid::config::Flag<double> smpi_wtime_sleep(
74 "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0);
75 static simgrid::config::Flag<double> smpi_init_sleep(
76 "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
78 void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t) = &smpi_comm_copy_buffer_callback;
82 int smpi_process_count()
87 simgrid::smpi::Process* smpi_process()
89 simgrid::MsgActorExt* msgExt = static_cast<simgrid::MsgActorExt*>(SIMIX_process_self()->data);
90 return static_cast<simgrid::smpi::Process*>(msgExt->data);
93 simgrid::smpi::Process* smpi_process_remote(int index)
95 return process_data[index_to_process_data[index]];
98 MPI_Comm smpi_process_comm_self(){
99 return smpi_process()->comm_self();
102 void smpi_process_init(int *argc, char ***argv){
103 simgrid::smpi::Process::init(argc, argv);
106 int smpi_process_index(){
107 return smpi_process()->index();
111 int smpi_global_size()
113 char *value = getenv("SMPI_GLOBAL_SIZE");
114 xbt_assert(value,"Please set env var SMPI_GLOBAL_SIZE to the expected number of processes.");
116 return xbt_str_parse_int(value, "SMPI_GLOBAL_SIZE contains a non-numerical value: %s");
119 void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
121 smpi_comm_copy_data_callback = callback;
124 void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size)
127 simgrid::kernel::activity::Comm *comm = dynamic_cast<simgrid::kernel::activity::Comm*>(synchro);
129 XBT_DEBUG("Copy the data over");
130 if(smpi_is_shared(buff)){
131 XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff);
132 }else if(smpi_is_shared((char*)comm->dst_buff)){
133 XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff);
136 if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && (static_cast<char*>(buff) >= smpi_start_data_exe)
137 && (static_cast<char*>(buff) < smpi_start_data_exe + smpi_size_data_exe )
139 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
141 smpi_switch_data_segment(
142 (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->src_proc->data)->data))->index()));
143 tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
144 memcpy(tmpbuff, buff, buff_size);
147 if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_start_data_exe)
148 && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){
149 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
150 smpi_switch_data_segment(
151 (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->dst_proc->data)->data))->index()));
154 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
155 memcpy(comm->dst_buff, tmpbuff, buff_size);
157 if (comm->detached) {
158 // if this is a detached send, the source buffer was duplicated by SMPI
159 // sender to make the original buffer available to the application ASAP
161 //It seems that the request is used after the call there this should be free somewhere else but where???
162 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
163 comm->src_buff = nullptr;
165 if(tmpbuff!=buff)xbt_free(tmpbuff);
170 void smpi_comm_null_copy_buffer_callback(smx_activity_t comm, void *buff, size_t buff_size)
172 /* nothing done in this version */
175 static void smpi_check_options(){
176 //check correctness of MPI parameters
178 xbt_assert(xbt_cfg_get_int("smpi/async-small-thresh") <= xbt_cfg_get_int("smpi/send-is-detached-thresh"));
180 if (xbt_cfg_is_default_value("smpi/host-speed")) {
181 XBT_INFO("You did not set the power of the host running the simulation. "
182 "The timings will certainly not be accurate. "
183 "Use the option \"--cfg=smpi/host-speed:<flops>\" to set its value."
184 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
187 xbt_assert(xbt_cfg_get_double("smpi/cpu-threshold") >=0,
188 "The 'smpi/cpu-threshold' option cannot have negative values [anymore]. If you want to discard "
189 "the simulation of any computation, please use 'smpi/simulate-computation:no' instead.");
193 return process_data != nullptr;
196 void smpi_global_init()
200 if (!MC_is_active()) {
201 global_timer = xbt_os_timer_new();
202 xbt_os_walltimer_start(global_timer);
205 if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') {
206 std::string filename {xbt_cfg_get_string("smpi/comp-adjustment-file")};
207 std::ifstream fstream(filename);
208 if (!fstream.is_open()) {
209 xbt_die("Could not open file %s. Does it exist?", filename.c_str());
213 typedef boost::tokenizer< boost::escaped_list_separator<char>> Tokenizer;
214 std::getline(fstream, line); // Skip the header line
215 while (std::getline(fstream, line)) {
217 Tokenizer::iterator it = tok.begin();
218 Tokenizer::iterator end = std::next(tok.begin());
220 std::string location = *it;
221 boost::trim(location);
222 location2speedup.insert(std::pair<std::string, double>(location, std::stod(*end)));
227 // This map holds for each computation unit (such as "default" or "process1" etc.)
228 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
229 // and the (computed) event_set.
230 std::map</* computation unit name */ std::string, papi_process_data> units2papi_setup;
232 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
233 if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
234 XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?"
235 " Expected version is %i",
238 typedef boost::tokenizer<boost::char_separator<char>> Tokenizer;
239 boost::char_separator<char> separator_units(";");
240 std::string str = std::string(xbt_cfg_get_string("smpi/papi-events"));
241 Tokenizer tokens(str, separator_units);
243 // Iterate over all the computational units. This could be
244 // processes, hosts, threads, ranks... You name it. I'm not exactly
245 // sure what we will support eventually, so I'll leave it at the
246 // general term "units".
247 for (auto& unit_it : tokens) {
248 boost::char_separator<char> separator_events(":");
249 Tokenizer event_tokens(unit_it, separator_events);
251 int event_set = PAPI_NULL;
252 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
253 // TODO: Should this let the whole simulation die?
254 XBT_CRITICAL("Could not create PAPI event set during init.");
257 // NOTE: We cannot use a map here, as we must obey the order of the counters
258 // This is important for PAPI: We need to map the values of counters back
259 // to the event_names (so, when PAPI_read() has finished)!
260 papi_counter_t counters2values;
262 // Iterate over all counters that were specified for this specific
264 // Note that we need to remove the name of the unit
265 // (that could also be the "default" value), which always comes first.
266 // Hence, we start at ++(events.begin())!
267 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); events_it++) {
269 int event_code = PAPI_NULL;
270 char* event_name = const_cast<char*>((*events_it).c_str());
271 if (PAPI_event_name_to_code(event_name, &event_code) == PAPI_OK) {
272 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
273 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
276 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
279 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
283 counters2values.push_back(
284 // We cannot just pass *events_it, as this is of type const basic_string
285 std::make_pair<std::string, long long>(std::string(*events_it), 0));
288 std::string unit_name = *(event_tokens.begin());
289 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
291 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
297 msg_bar_t finalization_barrier = nullptr;
298 if (process_count == 0){
299 process_count = SIMIX_process_count();
301 finalization_barrier = MSG_barrier_init(process_count);
303 smpi_universe_size = process_count;
304 process_data = new simgrid::smpi::Process*[process_count];
305 for (int i = 0; i < process_count; i++) {
306 process_data[i] = new simgrid::smpi::Process(i, finalization_barrier);
308 //if the process was launched through smpirun script we generate a global mpi_comm_world
309 //if not, we let MPI_COMM_NULL, and the comm world will be private to each mpi instance
311 group = new simgrid::smpi::Group(process_count);
312 MPI_COMM_WORLD = new simgrid::smpi::Comm(group, nullptr);
313 MPI_Attr_put(MPI_COMM_WORLD, MPI_UNIVERSE_SIZE, reinterpret_cast<void *>(process_count));
315 for (int i = 0; i < process_count; i++)
316 group->set_mapping(i, i);
320 void smpi_global_destroy()
322 int count = smpi_process_count();
324 smpi_bench_destroy();
325 smpi_shared_destroy();
326 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
327 delete MPI_COMM_WORLD->group();
328 MSG_barrier_destroy(process_data[0]->finalization_barrier());
330 smpi_deployment_cleanup_instances();
332 for (int i = 0; i < count; i++) {
333 if(process_data[i]->comm_self()!=MPI_COMM_NULL){
334 simgrid::smpi::Comm::destroy(process_data[i]->comm_self());
336 if(process_data[i]->comm_intra()!=MPI_COMM_NULL){
337 simgrid::smpi::Comm::destroy(process_data[i]->comm_intra());
339 xbt_os_timer_free(process_data[i]->timer());
340 xbt_mutex_destroy(process_data[i]->mailboxes_mutex());
341 delete process_data[i];
343 delete[] process_data;
344 process_data = nullptr;
346 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
347 MPI_COMM_WORLD->cleanup_smp();
348 MPI_COMM_WORLD->cleanup_attr<simgrid::smpi::Comm>();
349 if(simgrid::smpi::Colls::smpi_coll_cleanup_callback!=nullptr)
350 simgrid::smpi::Colls::smpi_coll_cleanup_callback();
351 delete MPI_COMM_WORLD;
354 MPI_COMM_WORLD = MPI_COMM_NULL;
356 if (!MC_is_active()) {
357 xbt_os_timer_free(global_timer);
360 xbt_free(index_to_process_data);
361 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
362 smpi_destroy_global_memory_segments();
368 static void smpi_init_logs(){
370 /* Connect log categories. See xbt/log.c */
372 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this function: xbt_log_appender_file.c depends on it
373 DO NOT connect this in XBT or so, or it will be useless to xbt_log_appender_file.c */
374 XBT_LOG_CONNECT(instr_smpi);
375 XBT_LOG_CONNECT(smpi_bench);
376 XBT_LOG_CONNECT(smpi_coll);
377 XBT_LOG_CONNECT(smpi_colls);
378 XBT_LOG_CONNECT(smpi_comm);
379 XBT_LOG_CONNECT(smpi_datatype);
380 XBT_LOG_CONNECT(smpi_dvfs);
381 XBT_LOG_CONNECT(smpi_group);
382 XBT_LOG_CONNECT(smpi_kernel);
383 XBT_LOG_CONNECT(smpi_mpi);
384 XBT_LOG_CONNECT(smpi_memory);
385 XBT_LOG_CONNECT(smpi_op);
386 XBT_LOG_CONNECT(smpi_pmpi);
387 XBT_LOG_CONNECT(smpi_request);
388 XBT_LOG_CONNECT(smpi_replay);
389 XBT_LOG_CONNECT(smpi_rma);
390 XBT_LOG_CONNECT(smpi_shared);
391 XBT_LOG_CONNECT(smpi_utils);
395 static void smpi_init_options(){
397 simgrid::smpi::Colls::set_collectives();
398 simgrid::smpi::Colls::smpi_coll_cleanup_callback=nullptr;
399 smpi_cpu_threshold = xbt_cfg_get_double("smpi/cpu-threshold");
400 smpi_host_speed = xbt_cfg_get_double("smpi/host-speed");
401 const char* smpi_privatize_option = xbt_cfg_get_string("smpi/privatize-global-variables");
402 if (std::strcmp(smpi_privatize_option, "no") == 0)
403 smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE;
404 else if (std::strcmp(smpi_privatize_option, "yes") == 0)
405 smpi_privatize_global_variables = SMPI_PRIVATIZE_DEFAULT;
406 else if (std::strcmp(smpi_privatize_option, "mmap") == 0)
407 smpi_privatize_global_variables = SMPI_PRIVATIZE_MMAP;
408 else if (std::strcmp(smpi_privatize_option, "dlopen") == 0)
409 smpi_privatize_global_variables = SMPI_PRIVATIZE_DLOPEN;
411 // Some compatibility stuff:
412 else if (std::strcmp(smpi_privatize_option, "1") == 0)
413 smpi_privatize_global_variables = SMPI_PRIVATIZE_DEFAULT;
414 else if (std::strcmp(smpi_privatize_option, "0") == 0)
415 smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE;
418 xbt_die("Invalid value for smpi/privatize-global-variables: %s",
419 smpi_privatize_option);
421 if (smpi_cpu_threshold < 0)
422 smpi_cpu_threshold = DBL_MAX;
424 char* val = xbt_cfg_get_string("smpi/shared-malloc");
425 if (!strcasecmp(val, "yes") || !strcmp(val, "1") || !strcasecmp(val, "on") || !strcasecmp(val, "global")) {
426 smpi_cfg_shared_malloc = shmalloc_global;
427 } else if (!strcasecmp(val, "local")) {
428 smpi_cfg_shared_malloc = shmalloc_local;
429 } else if (!strcasecmp(val, "no") || !strcmp(val, "0") || !strcasecmp(val, "off")) {
430 smpi_cfg_shared_malloc = shmalloc_none;
432 xbt_die("Invalid value '%s' for option smpi/shared-malloc. Possible values: 'on' or 'global', 'local', 'off'",
437 static int execute_command(const char * const argv[])
441 if (posix_spawnp(&pid, argv[0], nullptr, nullptr, (char* const*) argv, environ) != 0)
443 if (waitpid(pid, &status, 0) != pid)
448 typedef std::function<int(int argc, char *argv[])> smpi_entry_point_type;
449 typedef int (* smpi_c_entry_point_type)(int argc, char **argv);
450 typedef void (* smpi_fortran_entry_point_type)(void);
452 static int smpi_run_entry_point(smpi_entry_point_type entry_point, std::vector<std::string> args)
454 const int argc = args.size();
455 std::unique_ptr<char*[]> argv(new char*[argc + 1]);
456 for (int i = 0; i != argc; ++i)
457 argv[i] = args[i].empty() ? const_cast<char*>(""): &args[i].front();
458 argv[argc] = nullptr;
460 int res = entry_point(argc, argv.get());
462 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
463 smpi_process()->set_return_value(res);
468 // TODO, remove the number of functions involved here
469 static smpi_entry_point_type smpi_resolve_function(void* handle)
471 smpi_fortran_entry_point_type entry_point2 =
472 (smpi_fortran_entry_point_type) dlsym(handle, "user_main_");
473 if (entry_point2 != nullptr) {
474 // fprintf(stderr, "EP user_main_=%p\n", entry_point2);
475 return [entry_point2](int argc, char** argv) {
476 smpi_process_init(&argc, &argv);
482 smpi_c_entry_point_type entry_point = (smpi_c_entry_point_type) dlsym(handle, "main");
483 if (entry_point != nullptr) {
484 // fprintf(stderr, "EP main=%p\n", entry_point);
488 return smpi_entry_point_type();
491 int smpi_main(const char* executable, int argc, char *argv[])
493 srand(SMPI_RAND_SEED);
495 if (getenv("SMPI_PRETEND_CC") != nullptr) {
496 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
497 * configuration tools */
502 TRACE_global_init(&argc, argv);
503 TRACE_add_start_function(TRACE_smpi_alloc);
504 TRACE_add_end_function(TRACE_smpi_release);
506 SIMIX_global_init(&argc, argv);
507 MSG_init(&argc,argv);
509 SMPI_switch_data_segment = &smpi_switch_data_segment;
513 // parse the platform file: get the host list
514 SIMIX_create_environment(argv[1]);
515 SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback);
517 static std::size_t rank = 0;
519 if (smpi_privatize_global_variables == SMPI_PRIVATIZE_DLOPEN) {
521 std::string executable_copy = executable;
522 simix_global->default_function = [executable_copy](std::vector<std::string> args) {
523 return std::function<void()>([executable_copy, args] {
525 // Copy the dynamic library:
526 std::string target_executable = executable_copy
527 + "_" + std::to_string(getpid())
528 + "_" + std::to_string(rank++) + ".so";
529 // TODO, execute directly instead of relying on cp
530 const char* command1 [] = {
531 "cp", "--reflink=auto", "--", executable_copy.c_str(), target_executable.c_str(),
534 const char* command2 [] = {
535 "cp", "--", executable_copy.c_str(), target_executable.c_str(),
538 if (execute_command(command1) != 0 && execute_command(command2) != 0)
539 xbt_die("copy failed");
541 // Load the copy and resolve the entry point:
542 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);
543 unlink(target_executable.c_str());
544 if (handle == nullptr)
545 xbt_die("dlopen failed");
546 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
548 xbt_die("Could not resolve entry point");
550 smpi_run_entry_point(entry_point, args);
557 // Load the dynamic library and resolve the entry point:
558 void* handle = dlopen(executable, RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);
559 if (handle == nullptr)
560 xbt_die("dlopen failed for %s", executable);
561 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
563 xbt_die("main not found in %s", executable);
564 // TODO, register the executable for SMPI privatization
566 // Execute the same entry point for each simulated process:
567 simix_global->default_function = [entry_point](std::vector<std::string> args) {
568 return std::function<void()>([entry_point, args] {
569 smpi_run_entry_point(entry_point, args);
575 SIMIX_launch_application(argv[2]);
579 smpi_check_options();
581 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
582 smpi_initialize_global_memory_segments();
584 /* Clean IO before the run */
588 if (MC_is_active()) {
594 xbt_os_walltimer_stop(global_timer);
595 if (xbt_cfg_get_boolean("smpi/display-timing")){
596 double global_time = xbt_os_timer_elapsed(global_timer);
597 XBT_INFO("Simulated time: %g seconds. \n\n"
598 "The simulation took %g seconds (after parsing and platform setup)\n"
599 "%g seconds were actual computation of the application",
600 SIMIX_get_clock(), global_time , smpi_total_benched_time);
602 if (smpi_total_benched_time/global_time>=0.75)
603 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
604 "You may want to use sampling functions or trace replay to reduce this.");
607 int count = smpi_process_count();
609 for (i = 0; i < count; i++) {
610 if(process_data[i]->return_value()!=0){
611 ret=process_data[i]->return_value();//return first non 0 value
615 smpi_global_destroy();
622 // This function can be called from extern file, to initialize logs, options, and processes of smpi
623 // without the need of smpirun
628 smpi_check_options();
629 if (TRACE_is_enabled() && TRACE_is_configured())
631 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
632 smpi_initialize_global_memory_segments();
635 void SMPI_finalize(){
636 smpi_global_destroy();
639 void smpi_mpi_init() {
640 if(smpi_init_sleep > 0)
641 simcall_process_sleep(smpi_init_sleep);
644 double smpi_mpi_wtime(){
646 if (smpi_process()->initialized() != 0 && smpi_process()->finalized() == 0 && smpi_process()->sampling() == 0) {
648 time = SIMIX_get_clock();
649 // to avoid deadlocks if used as a break condition, such as
650 // while (MPI_Wtime(...) < time_limit) {
653 // because the time will not normally advance when only calls to MPI_Wtime
654 // are made -> deadlock (MPI_Wtime never reaches the time limit)
655 if(smpi_wtime_sleep > 0)
656 simcall_process_sleep(smpi_wtime_sleep);
659 time = SIMIX_get_clock();