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 "src/surf/surf_interface.hpp"
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 void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size)
128 simgrid::kernel::activity::Comm *comm = dynamic_cast<simgrid::kernel::activity::Comm*>(synchro);
130 XBT_DEBUG("Copy the data over");
131 if(smpi_is_shared(buff)){
132 XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff);
133 }else if(smpi_is_shared((char*)comm->dst_buff)){
134 XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff);
137 if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && (static_cast<char*>(buff) >= smpi_start_data_exe)
138 && (static_cast<char*>(buff) < smpi_start_data_exe + smpi_size_data_exe )
140 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
142 smpi_switch_data_segment(
143 (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->src_proc->data)->data))->index()));
144 tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
145 memcpy(tmpbuff, buff, buff_size);
148 if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_start_data_exe)
149 && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){
150 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
151 smpi_switch_data_segment(
152 (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->dst_proc->data)->data))->index()));
155 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
156 memcpy(comm->dst_buff, tmpbuff, buff_size);
158 if (comm->detached) {
159 // if this is a detached send, the source buffer was duplicated by SMPI
160 // sender to make the original buffer available to the application ASAP
162 //It seems that the request is used after the call there this should be free somewhere else but where???
163 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
164 comm->src_buff = nullptr;
166 if(tmpbuff!=buff)xbt_free(tmpbuff);
171 void smpi_comm_null_copy_buffer_callback(smx_activity_t comm, void *buff, size_t buff_size)
173 /* nothing done in this version */
176 static void smpi_check_options(){
177 //check correctness of MPI parameters
179 xbt_assert(xbt_cfg_get_int("smpi/async-small-thresh") <= xbt_cfg_get_int("smpi/send-is-detached-thresh"));
181 if (xbt_cfg_is_default_value("smpi/host-speed")) {
182 XBT_INFO("You did not set the power of the host running the simulation. "
183 "The timings will certainly not be accurate. "
184 "Use the option \"--cfg=smpi/host-speed:<flops>\" to set its value."
185 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
188 xbt_assert(xbt_cfg_get_double("smpi/cpu-threshold") >=0,
189 "The 'smpi/cpu-threshold' option cannot have negative values [anymore]. If you want to discard "
190 "the simulation of any computation, please use 'smpi/simulate-computation:no' instead.");
194 return process_data != nullptr;
197 void smpi_global_init()
201 if (!MC_is_active()) {
202 global_timer = xbt_os_timer_new();
203 xbt_os_walltimer_start(global_timer);
206 if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') {
207 std::string filename {xbt_cfg_get_string("smpi/comp-adjustment-file")};
208 std::ifstream fstream(filename);
209 if (!fstream.is_open()) {
210 xbt_die("Could not open file %s. Does it exist?", filename.c_str());
214 typedef boost::tokenizer< boost::escaped_list_separator<char>> Tokenizer;
215 std::getline(fstream, line); // Skip the header line
216 while (std::getline(fstream, line)) {
218 Tokenizer::iterator it = tok.begin();
219 Tokenizer::iterator end = std::next(tok.begin());
221 std::string location = *it;
222 boost::trim(location);
223 location2speedup.insert(std::pair<std::string, double>(location, std::stod(*end)));
228 // This map holds for each computation unit (such as "default" or "process1" etc.)
229 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
230 // and the (computed) event_set.
231 std::map</* computation unit name */ std::string, papi_process_data> units2papi_setup;
233 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
234 if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
235 XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?"
236 " Expected version is %i",
239 typedef boost::tokenizer<boost::char_separator<char>> Tokenizer;
240 boost::char_separator<char> separator_units(";");
241 std::string str = std::string(xbt_cfg_get_string("smpi/papi-events"));
242 Tokenizer tokens(str, separator_units);
244 // Iterate over all the computational units. This could be
245 // processes, hosts, threads, ranks... You name it. I'm not exactly
246 // sure what we will support eventually, so I'll leave it at the
247 // general term "units".
248 for (auto& unit_it : tokens) {
249 boost::char_separator<char> separator_events(":");
250 Tokenizer event_tokens(unit_it, separator_events);
252 int event_set = PAPI_NULL;
253 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
254 // TODO: Should this let the whole simulation die?
255 XBT_CRITICAL("Could not create PAPI event set during init.");
258 // NOTE: We cannot use a map here, as we must obey the order of the counters
259 // This is important for PAPI: We need to map the values of counters back
260 // to the event_names (so, when PAPI_read() has finished)!
261 papi_counter_t counters2values;
263 // Iterate over all counters that were specified for this specific
265 // Note that we need to remove the name of the unit
266 // (that could also be the "default" value), which always comes first.
267 // Hence, we start at ++(events.begin())!
268 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); events_it++) {
270 int event_code = PAPI_NULL;
271 char* event_name = const_cast<char*>((*events_it).c_str());
272 if (PAPI_event_name_to_code(event_name, &event_code) == PAPI_OK) {
273 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
274 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
277 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
280 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
284 counters2values.push_back(
285 // We cannot just pass *events_it, as this is of type const basic_string
286 std::make_pair<std::string, long long>(std::string(*events_it), 0));
289 std::string unit_name = *(event_tokens.begin());
290 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
292 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
298 msg_bar_t finalization_barrier = nullptr;
299 if (process_count == 0){
300 process_count = SIMIX_process_count();
302 finalization_barrier = MSG_barrier_init(process_count);
304 smpi_universe_size = process_count;
305 process_data = new simgrid::smpi::Process*[process_count];
306 for (int i = 0; i < process_count; i++) {
307 process_data[i] = new simgrid::smpi::Process(i, finalization_barrier);
309 //if the process was launched through smpirun script we generate a global mpi_comm_world
310 //if not, we let MPI_COMM_NULL, and the comm world will be private to each mpi instance
312 group = new simgrid::smpi::Group(process_count);
313 MPI_COMM_WORLD = new simgrid::smpi::Comm(group, nullptr);
314 MPI_Attr_put(MPI_COMM_WORLD, MPI_UNIVERSE_SIZE, reinterpret_cast<void *>(process_count));
316 for (int i = 0; i < process_count; i++)
317 group->set_mapping(i, i);
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 */
502 TRACE_global_init(&argc, argv);
504 SIMIX_global_init(&argc, argv);
505 MSG_init(&argc,argv);
507 SMPI_switch_data_segment = &smpi_switch_data_segment;
509 // parse the platform file: get the host list
510 SIMIX_create_environment(argv[1]);
511 SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback);
513 static std::size_t rank = 0;
515 if (smpi_privatize_global_variables == SMPI_PRIVATIZE_DLOPEN) {
517 std::string executable_copy = executable;
518 simix_global->default_function = [executable_copy](std::vector<std::string> args) {
519 return std::function<void()>([executable_copy, args] {
521 // Copy the dynamic library:
522 std::string target_executable = executable_copy
523 + "_" + std::to_string(getpid())
524 + "_" + std::to_string(rank++) + ".so";
525 // TODO, execute directly instead of relying on cp
526 const char* command1 [] = {
527 "cp", "--reflink=auto", "--", executable_copy.c_str(), target_executable.c_str(),
530 const char* command2 [] = {
531 "cp", "--", executable_copy.c_str(), target_executable.c_str(),
534 if (execute_command(command1) != 0 && execute_command(command2) != 0)
535 xbt_die("copy failed");
537 // Load the copy and resolve the entry point:
538 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);
539 unlink(target_executable.c_str());
540 if (handle == nullptr)
541 xbt_die("dlopen failed");
542 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
544 xbt_die("Could not resolve entry point");
546 smpi_run_entry_point(entry_point, args);
553 // Load the dynamic library and resolve the entry point:
554 void* handle = dlopen(executable, RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);
555 if (handle == nullptr)
556 xbt_die("dlopen failed for %s", executable);
557 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
559 xbt_die("main not found in %s", executable);
560 // TODO, register the executable for SMPI privatization
562 // Execute the same entry point for each simulated process:
563 simix_global->default_function = [entry_point](std::vector<std::string> args) {
564 return std::function<void()>([entry_point, args] {
565 smpi_run_entry_point(entry_point, args);
571 SIMIX_launch_application(argv[2]);
575 /* Clean IO before the run */
579 if (MC_is_active()) {
585 xbt_os_walltimer_stop(global_timer);
586 if (xbt_cfg_get_boolean("smpi/display-timing")){
587 double global_time = xbt_os_timer_elapsed(global_timer);
588 XBT_INFO("Simulated time: %g seconds. \n\n"
589 "The simulation took %g seconds (after parsing and platform setup)\n"
590 "%g seconds were actual computation of the application",
591 SIMIX_get_clock(), global_time , smpi_total_benched_time);
593 if (smpi_total_benched_time/global_time>=0.75)
594 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
595 "You may want to use sampling functions or trace replay to reduce this.");
598 int count = smpi_process_count();
600 for (i = 0; i < count; i++) {
601 if(process_data[i]->return_value()!=0){
602 ret=process_data[i]->return_value();//return first non 0 value
606 smpi_global_destroy();
613 // Called either directly from the user code, or from the code called by smpirun
618 smpi_check_options();
620 simgrid::surf::surfExitCallbacks.connect(TRACE_smpi_release);
621 if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
622 smpi_initialize_global_memory_segments();
625 void SMPI_finalize(){
626 smpi_global_destroy();
629 void smpi_mpi_init() {
630 if(smpi_init_sleep > 0)
631 simcall_process_sleep(smpi_init_sleep);
634 double smpi_mpi_wtime(){
636 if (smpi_process()->initialized() != 0 && smpi_process()->finalized() == 0 && smpi_process()->sampling() == 0) {
638 time = SIMIX_get_clock();
639 // to avoid deadlocks if used as a break condition, such as
640 // while (MPI_Wtime(...) < time_limit) {
643 // because the time will not normally advance when only calls to MPI_Wtime
644 // are made -> deadlock (MPI_Wtime never reaches the time limit)
645 if(smpi_wtime_sleep > 0)
646 simcall_process_sleep(smpi_wtime_sleep);
649 time = SIMIX_get_clock();