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 if (process_count == 0){
298 process_count = SIMIX_process_count();
301 smpi_universe_size = process_count;
302 process_data = new simgrid::smpi::Process*[process_count];
303 for (int i = 0; i < process_count; i++) {
304 process_data[i] = new simgrid::smpi::Process(i);
306 //if the process was launched through smpirun script we generate a global mpi_comm_world
307 //if not, we let MPI_COMM_NULL, and the comm world will be private to each mpi instance
309 group = new simgrid::smpi::Group(process_count);
310 MPI_COMM_WORLD = new simgrid::smpi::Comm(group, nullptr);
311 MPI_Attr_put(MPI_COMM_WORLD, MPI_UNIVERSE_SIZE, reinterpret_cast<void *>(process_count));
312 msg_bar_t bar = MSG_barrier_init(process_count);
314 for (int i = 0; i < process_count; i++) {
315 group->set_mapping(i, i);
316 process_data[i]->set_finalization_barrier(bar);
321 void smpi_global_destroy()
323 int count = smpi_process_count();
325 smpi_bench_destroy();
326 smpi_shared_destroy();
327 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
328 delete MPI_COMM_WORLD->group();
329 MSG_barrier_destroy(process_data[0]->finalization_barrier());
331 smpi_deployment_cleanup_instances();
333 for (int i = 0; i < count; i++) {
334 if(process_data[i]->comm_self()!=MPI_COMM_NULL){
335 simgrid::smpi::Comm::destroy(process_data[i]->comm_self());
337 if(process_data[i]->comm_intra()!=MPI_COMM_NULL){
338 simgrid::smpi::Comm::destroy(process_data[i]->comm_intra());
340 xbt_os_timer_free(process_data[i]->timer());
341 xbt_mutex_destroy(process_data[i]->mailboxes_mutex());
342 delete process_data[i];
344 delete[] process_data;
345 process_data = nullptr;
347 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
348 MPI_COMM_WORLD->cleanup_smp();
349 MPI_COMM_WORLD->cleanup_attr<simgrid::smpi::Comm>();
350 if(simgrid::smpi::Colls::smpi_coll_cleanup_callback!=nullptr)
351 simgrid::smpi::Colls::smpi_coll_cleanup_callback();
352 delete MPI_COMM_WORLD;
355 MPI_COMM_WORLD = MPI_COMM_NULL;
357 if (!MC_is_active()) {
358 xbt_os_timer_free(global_timer);
361 xbt_free(index_to_process_data);
362 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
363 smpi_destroy_global_memory_segments();
369 static void smpi_init_logs(){
371 /* Connect log categories. See xbt/log.c */
373 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this function: xbt_log_appender_file.c depends on it
374 DO NOT connect this in XBT or so, or it will be useless to xbt_log_appender_file.c */
375 XBT_LOG_CONNECT(instr_smpi);
376 XBT_LOG_CONNECT(smpi_bench);
377 XBT_LOG_CONNECT(smpi_coll);
378 XBT_LOG_CONNECT(smpi_colls);
379 XBT_LOG_CONNECT(smpi_comm);
380 XBT_LOG_CONNECT(smpi_datatype);
381 XBT_LOG_CONNECT(smpi_dvfs);
382 XBT_LOG_CONNECT(smpi_group);
383 XBT_LOG_CONNECT(smpi_kernel);
384 XBT_LOG_CONNECT(smpi_mpi);
385 XBT_LOG_CONNECT(smpi_memory);
386 XBT_LOG_CONNECT(smpi_op);
387 XBT_LOG_CONNECT(smpi_pmpi);
388 XBT_LOG_CONNECT(smpi_request);
389 XBT_LOG_CONNECT(smpi_replay);
390 XBT_LOG_CONNECT(smpi_rma);
391 XBT_LOG_CONNECT(smpi_shared);
392 XBT_LOG_CONNECT(smpi_utils);
396 static void smpi_init_options(){
398 simgrid::smpi::Colls::set_collectives();
399 simgrid::smpi::Colls::smpi_coll_cleanup_callback=nullptr;
400 smpi_cpu_threshold = xbt_cfg_get_double("smpi/cpu-threshold");
401 smpi_host_speed = xbt_cfg_get_double("smpi/host-speed");
402 const char* smpi_privatize_option = xbt_cfg_get_string("smpi/privatize-global-variables");
403 if (std::strcmp(smpi_privatize_option, "no") == 0)
404 smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE;
405 else if (std::strcmp(smpi_privatize_option, "yes") == 0)
406 smpi_privatize_global_variables = SMPI_PRIVATIZE_DEFAULT;
407 else if (std::strcmp(smpi_privatize_option, "mmap") == 0)
408 smpi_privatize_global_variables = SMPI_PRIVATIZE_MMAP;
409 else if (std::strcmp(smpi_privatize_option, "dlopen") == 0)
410 smpi_privatize_global_variables = SMPI_PRIVATIZE_DLOPEN;
412 // Some compatibility stuff:
413 else if (std::strcmp(smpi_privatize_option, "1") == 0)
414 smpi_privatize_global_variables = SMPI_PRIVATIZE_DEFAULT;
415 else if (std::strcmp(smpi_privatize_option, "0") == 0)
416 smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE;
419 xbt_die("Invalid value for smpi/privatize-global-variables: %s",
420 smpi_privatize_option);
422 if (smpi_cpu_threshold < 0)
423 smpi_cpu_threshold = DBL_MAX;
425 char* val = xbt_cfg_get_string("smpi/shared-malloc");
426 if (!strcasecmp(val, "yes") || !strcmp(val, "1") || !strcasecmp(val, "on") || !strcasecmp(val, "global")) {
427 smpi_cfg_shared_malloc = shmalloc_global;
428 } else if (!strcasecmp(val, "local")) {
429 smpi_cfg_shared_malloc = shmalloc_local;
430 } else if (!strcasecmp(val, "no") || !strcmp(val, "0") || !strcasecmp(val, "off")) {
431 smpi_cfg_shared_malloc = shmalloc_none;
433 xbt_die("Invalid value '%s' for option smpi/shared-malloc. Possible values: 'on' or 'global', 'local', 'off'",
438 static int execute_command(const char * const argv[])
442 if (posix_spawnp(&pid, argv[0], nullptr, nullptr, (char* const*) argv, environ) != 0)
444 if (waitpid(pid, &status, 0) != pid)
449 typedef std::function<int(int argc, char *argv[])> smpi_entry_point_type;
450 typedef int (* smpi_c_entry_point_type)(int argc, char **argv);
451 typedef void (* smpi_fortran_entry_point_type)(void);
453 static int smpi_run_entry_point(smpi_entry_point_type entry_point, std::vector<std::string> args)
455 const int argc = args.size();
456 std::unique_ptr<char*[]> argv(new char*[argc + 1]);
457 for (int i = 0; i != argc; ++i)
458 argv[i] = args[i].empty() ? const_cast<char*>(""): &args[i].front();
459 argv[argc] = nullptr;
461 int res = entry_point(argc, argv.get());
463 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
464 smpi_process()->set_return_value(res);
469 // TODO, remove the number of functions involved here
470 static smpi_entry_point_type smpi_resolve_function(void* handle)
472 smpi_fortran_entry_point_type entry_point2 =
473 (smpi_fortran_entry_point_type) dlsym(handle, "user_main_");
474 if (entry_point2 != nullptr) {
475 // fprintf(stderr, "EP user_main_=%p\n", entry_point2);
476 return [entry_point2](int argc, char** argv) {
477 smpi_process_init(&argc, &argv);
483 smpi_c_entry_point_type entry_point = (smpi_c_entry_point_type) dlsym(handle, "main");
484 if (entry_point != nullptr) {
485 // fprintf(stderr, "EP main=%p\n", entry_point);
489 return smpi_entry_point_type();
492 int smpi_main(const char* executable, int argc, char *argv[])
494 srand(SMPI_RAND_SEED);
496 if (getenv("SMPI_PRETEND_CC") != nullptr) {
497 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
498 * configuration tools */
503 TRACE_global_init(&argc, argv);
504 TRACE_add_start_function(TRACE_smpi_alloc);
505 TRACE_add_end_function(TRACE_smpi_release);
507 SIMIX_global_init(&argc, argv);
508 MSG_init(&argc,argv);
510 SMPI_switch_data_segment = &smpi_switch_data_segment;
514 // parse the platform file: get the host list
515 SIMIX_create_environment(argv[1]);
516 SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback);
518 static std::size_t rank = 0;
520 if (smpi_privatize_global_variables == SMPI_PRIVATIZE_DLOPEN) {
522 std::string executable_copy = executable;
523 simix_global->default_function = [executable_copy](std::vector<std::string> args) {
524 return std::function<void()>([executable_copy, args] {
526 // Copy the dynamic library:
527 std::string target_executable = executable_copy
528 + "_" + std::to_string(getpid())
529 + "_" + std::to_string(rank++) + ".so";
530 // TODO, execute directly instead of relying on cp
531 const char* command1 [] = {
532 "cp", "--reflink=auto", "--", executable_copy.c_str(), target_executable.c_str(),
535 const char* command2 [] = {
536 "cp", "--", executable_copy.c_str(), target_executable.c_str(),
539 if (execute_command(command1) != 0 && execute_command(command2) != 0)
540 xbt_die("copy failed");
542 // Load the copy and resolve the entry point:
543 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);
544 unlink(target_executable.c_str());
545 if (handle == nullptr)
546 xbt_die("dlopen failed");
547 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
549 xbt_die("Could not resolve entry point");
551 smpi_run_entry_point(entry_point, args);
558 // Load the dynamic library and resolve the entry point:
559 void* handle = dlopen(executable, RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);
560 if (handle == nullptr)
561 xbt_die("dlopen failed for %s", executable);
562 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
564 xbt_die("main not found in %s", executable);
565 // TODO, register the executable for SMPI privatization
567 // Execute the same entry point for each simulated process:
568 simix_global->default_function = [entry_point](std::vector<std::string> args) {
569 return std::function<void()>([entry_point, args] {
570 smpi_run_entry_point(entry_point, args);
576 SIMIX_launch_application(argv[2]);
580 smpi_check_options();
582 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
583 smpi_initialize_global_memory_segments();
585 /* Clean IO before the run */
589 if (MC_is_active()) {
595 xbt_os_walltimer_stop(global_timer);
596 if (xbt_cfg_get_boolean("smpi/display-timing")){
597 double global_time = xbt_os_timer_elapsed(global_timer);
598 XBT_INFO("Simulated time: %g seconds. \n\n"
599 "The simulation took %g seconds (after parsing and platform setup)\n"
600 "%g seconds were actual computation of the application",
601 SIMIX_get_clock(), global_time , smpi_total_benched_time);
603 if (smpi_total_benched_time/global_time>=0.75)
604 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
605 "You may want to use sampling functions or trace replay to reduce this.");
608 int count = smpi_process_count();
610 for (i = 0; i < count; i++) {
611 if(process_data[i]->return_value()!=0){
612 ret=process_data[i]->return_value();//return first non 0 value
616 smpi_global_destroy();
623 // This function can be called from extern file, to initialize logs, options, and processes of smpi
624 // without the need of smpirun
629 smpi_check_options();
630 if (TRACE_is_enabled() && TRACE_is_configured())
632 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
633 smpi_initialize_global_memory_segments();
636 void SMPI_finalize(){
637 smpi_global_destroy();
640 void smpi_mpi_init() {
641 if(smpi_init_sleep > 0)
642 simcall_process_sleep(smpi_init_sleep);
645 double smpi_mpi_wtime(){
647 if (smpi_process()->initialized() != 0 && smpi_process()->finalized() == 0 && smpi_process()->sampling() == 0) {
649 time = SIMIX_get_clock();
650 // to avoid deadlocks if used as a break condition, such as
651 // while (MPI_Wtime(...) < time_limit) {
654 // because the time will not normally advance when only calls to MPI_Wtime
655 // are made -> deadlock (MPI_Wtime never reaches the time limit)
656 if(smpi_wtime_sleep > 0)
657 simcall_process_sleep(smpi_wtime_sleep);
660 time = SIMIX_get_clock();