1 /* Copyright (c) 2007-2020. 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. */
7 #include "simgrid/s4u/Engine.hpp"
8 #include "simgrid/plugins/file_system.h"
9 #include "smpi_coll.hpp"
10 #include "smpi_f2c.hpp"
11 #include "smpi_host.hpp"
12 #include "smpi_config.hpp"
13 #include "src/kernel/activity/CommImpl.hpp"
14 #include "src/simix/smx_private.hpp"
15 #include "src/smpi/include/smpi_actor.hpp"
16 #include "xbt/config.hpp"
19 #include <boost/algorithm/string.hpp> /* split */
20 #include <boost/tokenizer.hpp>
22 #include <cstdint> /* intmax_t */
29 #include <sys/sendfile.h>
36 #if not defined(__APPLE__) && not defined(__HAIKU__)
40 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)");
43 extern "C" void for_rtl_init_ (int *, char **);
44 extern "C" void for_rtl_finish_ ();
46 extern "C" void __io_set_argc(int);
47 extern "C" void __io_set_argv(char **);
49 extern "C" void _gfortran_set_args(int, char **);
52 /* RTLD_DEEPBIND is a bad idea of GNU ld that obviously does not exist on other platforms
53 * See https://www.akkadia.org/drepper/dsohowto.pdf
54 * and https://lists.freebsd.org/pipermail/freebsd-current/2016-March/060284.html
56 #if !RTLD_DEEPBIND || HAVE_SANITIZER_ADDRESS || HAVE_SANITIZER_THREAD
57 #define WANT_RTLD_DEEPBIND 0
59 #define WANT_RTLD_DEEPBIND RTLD_DEEPBIND
63 std::string papi_default_config_name = "default";
64 std::map</* computation unit name */ std::string, papi_process_data> units2papi_setup;
67 std::unordered_map<std::string, double> location2speedup;
69 static int smpi_exit_status = 0;
70 extern double smpi_total_benched_time;
71 xbt_os_timer_t global_timer;
72 static std::vector<std::string> privatize_libs_paths;
74 * Setting MPI_COMM_WORLD to MPI_COMM_UNINITIALIZED (it's a variable)
75 * is important because the implementation of MPI_Comm checks
76 * "this == MPI_COMM_UNINITIALIZED"? If yes, it uses smpi_process()->comm_world()
78 * This is basically how we only have one global variable but all processes have
79 * different communicators (the one their SMPI instance uses).
81 * See smpi_comm.cpp and the functions therein for details.
83 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
84 // No instance gets manually created; check also the smpirun.in script as
85 // this default name is used there as well (when the <actor> tag is generated).
86 static const std::string smpi_default_instance_name("smpirun");
87 static simgrid::config::Flag<double> smpi_init_sleep(
88 "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
90 void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*,
91 size_t) = &smpi_comm_copy_buffer_callback;
93 simgrid::smpi::ActorExt* smpi_process()
95 simgrid::s4u::ActorPtr me = simgrid::s4u::Actor::self();
97 if (me == nullptr) // This happens sometimes (eg, when linking against NS3 because it pulls openMPI...)
100 return me->extension<simgrid::smpi::ActorExt>();
103 simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor)
105 if (actor.get() == nullptr)
107 return actor->extension<simgrid::smpi::ActorExt>();
110 MPI_Comm smpi_process_comm_self(){
111 return smpi_process()->comm_self();
114 MPI_Info smpi_process_info_env(){
115 return smpi_process()->info_env();
118 void * smpi_process_get_user_data(){
119 return simgrid::s4u::Actor::self()->get_data();
122 void smpi_process_set_user_data(void *data){
123 simgrid::s4u::Actor::self()->set_data(data);
126 void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
128 static void (*saved_callback)(smx_activity_t, void*, size_t);
129 saved_callback = callback;
130 smpi_comm_copy_data_callback = [](simgrid::kernel::activity::CommImpl* comm, void* buff, size_t size) {
131 saved_callback(comm, buff, size);
135 static void memcpy_private(void* dest, const void* src, const std::vector<std::pair<size_t, size_t>>& private_blocks)
137 for (auto const& block : private_blocks)
138 memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first);
141 static void check_blocks(const std::vector<std::pair<size_t, size_t>>& private_blocks, size_t buff_size)
143 for (auto const& block : private_blocks)
144 xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
147 static void smpi_cleanup_comm_after_copy(simgrid::kernel::activity::CommImpl* comm, void* buff){
148 if (comm->detached()) {
149 // if this is a detached send, the source buffer was duplicated by SMPI
150 // sender to make the original buffer available to the application ASAP
152 //It seems that the request is used after the call there this should be free somewhere else but where???
153 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
154 comm->src_buff_ = nullptr;
158 void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff, size_t buff_size)
160 size_t src_offset = 0;
161 size_t dst_offset = 0;
162 std::vector<std::pair<size_t, size_t>> src_private_blocks;
163 std::vector<std::pair<size_t, size_t>> dst_private_blocks;
164 XBT_DEBUG("Copy the data over");
165 if(smpi_is_shared(buff, src_private_blocks, &src_offset)) {
166 src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
167 if (src_private_blocks.size()==0){//simple shared malloc ... return.
168 XBT_VERB("Sender is shared. Let's ignore it.");
169 smpi_cleanup_comm_after_copy(comm, buff);
174 src_private_blocks.clear();
175 src_private_blocks.push_back(std::make_pair(0, buff_size));
177 if (smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset)) {
178 dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
179 if (dst_private_blocks.size()==0){//simple shared malloc ... return.
180 XBT_VERB("Receiver is shared. Let's ignore it.");
181 smpi_cleanup_comm_after_copy(comm, buff);
186 dst_private_blocks.clear();
187 dst_private_blocks.push_back(std::make_pair(0, buff_size));
189 check_blocks(src_private_blocks, buff_size);
190 check_blocks(dst_private_blocks, buff_size);
191 auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
192 check_blocks(private_blocks, buff_size);
194 if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) &&
195 (static_cast<char*>(buff) >= smpi_data_exe_start) &&
196 (static_cast<char*>(buff) < smpi_data_exe_start + smpi_data_exe_size)) {
197 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
198 smpi_switch_data_segment(comm->src_actor_->iface());
199 tmpbuff = xbt_malloc(buff_size);
200 memcpy_private(tmpbuff, buff, private_blocks);
203 if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) &&
204 ((char*)comm->dst_buff_ >= smpi_data_exe_start) &&
205 ((char*)comm->dst_buff_ < smpi_data_exe_start + smpi_data_exe_size)) {
206 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
207 smpi_switch_data_segment(comm->dst_actor_->iface());
209 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_);
210 memcpy_private(comm->dst_buff_, tmpbuff, private_blocks);
212 smpi_cleanup_comm_after_copy(comm,buff);
217 void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, void*, size_t)
219 /* nothing done in this version */
223 return MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED;
226 static void smpi_init_papi()
229 // This map holds for each computation unit (such as "default" or "process1" etc.)
230 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
231 // and the (computed) event_set.
233 if (not smpi_cfg_papi_events_file().empty()) {
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 %u", PAPI_VER_CURRENT);
238 typedef boost::tokenizer<boost::char_separator<char>> Tokenizer;
239 boost::char_separator<char> separator_units(";");
240 std::string str = smpi_cfg_papi_events_file();
241 Tokenizer tokens(str, separator_units);
243 // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it.
244 // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units".
245 for (auto const& unit_it : tokens) {
246 boost::char_separator<char> separator_events(":");
247 Tokenizer event_tokens(unit_it, separator_events);
249 int event_set = PAPI_NULL;
250 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
251 // TODO: Should this let the whole simulation die?
252 XBT_CRITICAL("Could not create PAPI event set during init.");
255 // NOTE: We cannot use a map here, as we must obey the order of the counters
256 // This is important for PAPI: We need to map the values of counters back to the event_names (so, when PAPI_read()
258 papi_counter_t counters2values;
260 // Iterate over all counters that were specified for this specific unit.
261 // Note that we need to remove the name of the unit (that could also be the "default" value), which always comes
262 // first. Hence, we start at ++(events.begin())!
263 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) {
264 int event_code = PAPI_NULL;
265 char* event_name = const_cast<char*>((*events_it).c_str());
266 if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) {
267 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
270 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
271 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
274 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
276 counters2values.push_back(
277 // We cannot just pass *events_it, as this is of type const basic_string
278 std::make_pair(std::string(*events_it), 0LL));
281 std::string unit_name = *(event_tokens.begin());
282 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
284 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
292 typedef std::function<int(int argc, char *argv[])> smpi_entry_point_type;
293 typedef int (* smpi_c_entry_point_type)(int argc, char **argv);
294 typedef void (*smpi_fortran_entry_point_type)();
296 template <typename F>
297 static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path, std::vector<std::string> args)
299 // copy C strings, we need them writable
300 std::vector<char*>* args4argv = new std::vector<char*>(args.size());
301 std::transform(begin(args), end(args), begin(*args4argv), [](const std::string& s) { return xbt_strdup(s.c_str()); });
303 // set argv[0] to executable_path
304 xbt_free((*args4argv)[0]);
305 (*args4argv)[0] = xbt_strdup(executable_path.c_str());
308 // take a copy of args4argv to keep reference of the allocated strings
309 const std::vector<char*> args2str(*args4argv);
311 int argc = args4argv->size();
312 args4argv->push_back(nullptr);
313 char** argv = args4argv->data();
316 for_rtl_init_ (&argc, argv);
321 _gfortran_set_args(argc, argv);
323 int res = entry_point(argc, argv);
328 for (char* s : args2str)
334 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
335 if (smpi_exit_status == 0)
336 smpi_exit_status = res;
342 // TODO, remove the number of functions involved here
343 static smpi_entry_point_type smpi_resolve_function(void* handle)
345 smpi_fortran_entry_point_type entry_point_fortran = (smpi_fortran_entry_point_type)dlsym(handle, "user_main_");
346 if (entry_point_fortran != nullptr) {
347 return [entry_point_fortran](int, char**) {
348 entry_point_fortran();
353 smpi_c_entry_point_type entry_point = (smpi_c_entry_point_type)dlsym(handle, "main");
354 if (entry_point != nullptr) {
358 return smpi_entry_point_type();
361 static void smpi_copy_file(const std::string& src, const std::string& target, off_t fdin_size)
363 int fdin = open(src.c_str(), O_RDONLY);
364 xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str());
365 int fdout = open(target.c_str(), O_CREAT | O_RDWR, S_IRWXU);
366 xbt_assert(fdout >= 0, "Cannot write into %s", target.c_str());
368 XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast<intmax_t>(fdin_size), target.c_str());
370 ssize_t sent_size = sendfile(fdout, fdin, NULL, fdin_size);
371 if (sent_size == fdin_size) {
375 } else if (sent_size != -1 || errno != ENOSYS) {
376 xbt_die("Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(),
377 sent_size, static_cast<intmax_t>(fdin_size), errno, strerror(errno));
380 // If this point is reached, sendfile() actually is not available. Copy file by hand.
381 const int bufsize = 1024 * 1024 * 4;
382 char* buf = new char[bufsize];
383 while (int got = read(fdin, buf, bufsize)) {
385 xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str());
389 while (int done = write(fdout, p, todo)) {
391 xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str());
404 #if not defined(__APPLE__) && not defined(__HAIKU__)
405 static int visit_libs(struct dl_phdr_info* info, size_t, void* data)
407 char* libname = (char*)(data);
408 const char *path = info->dlpi_name;
409 if(strstr(path, libname)){
410 strncpy(libname, path, 512);
418 static void smpi_init_privatization_dlopen(const std::string& executable)
420 // Prepare the copy of the binary (get its size)
421 struct stat fdin_stat;
422 stat(executable.c_str(), &fdin_stat);
423 off_t fdin_size = fdin_stat.st_size;
425 std::string libnames = simgrid::config::get_value<std::string>("smpi/privatize-libs");
426 if (not libnames.empty()) {
428 std::vector<std::string> privatize_libs;
429 boost::split(privatize_libs, libnames, boost::is_any_of(";"));
431 for (auto const& libname : privatize_libs) {
432 // load the library once to add it to the local libs, to get the absolute path
433 void* libhandle = dlopen(libname.c_str(), RTLD_LAZY);
434 // get library name from path
435 char fullpath[512] = {'\0'};
436 strncpy(fullpath, libname.c_str(), 511);
437 #if not defined(__APPLE__) && not defined(__HAIKU__)
438 xbt_assert(0 != dl_iterate_phdr(visit_libs, fullpath),
439 "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath);
440 XBT_DEBUG("Extra lib to privatize '%s' found", fullpath);
442 xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku");
444 privatize_libs_paths.push_back(fullpath);
449 simgrid::s4u::Engine::get_instance()->register_default([executable, fdin_size](std::vector<std::string> args) {
450 return std::function<void()>([executable, fdin_size, args] {
451 static std::size_t rank = 0;
452 // Copy the dynamic library:
453 std::string target_executable =
454 executable + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so";
456 smpi_copy_file(executable, target_executable, fdin_size);
457 // if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one.
458 std::vector<std::string> target_libs;
459 for (auto const& libpath : privatize_libs_paths) {
460 // if we were given a full path, strip it
461 size_t index = libpath.find_last_of("/\\");
463 if (index != std::string::npos)
464 libname = libpath.substr(index + 1);
466 if (not libname.empty()) {
467 // load the library to add it to the local libs, to get the absolute path
468 struct stat fdin_stat2;
469 stat(libpath.c_str(), &fdin_stat2);
470 off_t fdin_size2 = fdin_stat2.st_size;
472 // Copy the dynamic library, the new name must be the same length as the old one
473 // just replace the name with 7 digits for the rank and the rest of the name.
474 unsigned int pad = 7;
475 if (libname.length() < pad)
476 pad = libname.length();
477 std::string target_lib =
478 std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
479 target_libs.push_back(target_lib);
480 XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2);
481 smpi_copy_file(libpath, target_lib, fdin_size2);
483 std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_lib + "/g' " + target_executable;
484 xbt_assert(system(sedcommand.c_str()) == 0, "error while applying sed command %s \n", sedcommand.c_str());
489 // Load the copy and resolve the entry point:
490 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND);
491 int saved_errno = errno;
492 if (simgrid::config::get_value<bool>("smpi/keep-temps") == false) {
493 unlink(target_executable.c_str());
494 for (const std::string& target_lib : target_libs)
495 unlink(target_lib.c_str());
497 xbt_assert(handle != nullptr, "dlopen failed: %s (errno: %d -- %s)", dlerror(), saved_errno,
498 strerror(saved_errno));
500 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
501 xbt_assert(entry_point, "Could not resolve entry point");
502 smpi_run_entry_point(entry_point, executable, args);
507 static void smpi_init_privatization_no_dlopen(const std::string& executable)
509 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
510 smpi_prepare_global_memory_segment();
512 // Load the dynamic library and resolve the entry point:
513 void* handle = dlopen(executable.c_str(), RTLD_LAZY | RTLD_LOCAL);
514 xbt_assert(handle != nullptr, "dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno,
516 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
517 xbt_assert(entry_point, "main not found in %s", executable.c_str());
519 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
520 smpi_backup_global_memory_segment();
522 // Execute the same entry point for each simulated process:
523 simgrid::s4u::Engine::get_instance()->register_default([entry_point, executable](std::vector<std::string> args) {
524 return std::function<void()>(
525 [entry_point, executable, args] { smpi_run_entry_point(entry_point, executable, args); });
529 int smpi_main(const char* executable, int argc, char* argv[])
531 if (getenv("SMPI_PRETEND_CC") != nullptr) {
532 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
533 * configuration tools */
537 SMPI_switch_data_segment = &smpi_switch_data_segment;
539 simgrid::instr::init();
540 SIMIX_global_init(&argc, argv);
542 auto engine = simgrid::s4u::Engine::get_instance();
544 sg_storage_file_system_init();
545 // parse the platform file: get the host list
546 engine->load_platform(argv[1]);
547 SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback);
549 if (smpi_cfg_privatization() == SmpiPrivStrategies::DLOPEN)
550 smpi_init_privatization_dlopen(executable);
552 smpi_init_privatization_no_dlopen(executable);
554 simgrid::smpi::colls::set_collectives();
555 simgrid::smpi::colls::smpi_coll_cleanup_callback = nullptr;
559 /* This is a ... heavy way to count the MPI ranks */
561 simgrid::s4u::Actor::on_creation.connect([&rank_counts](const simgrid::s4u::Actor& actor) {
562 if (not actor.is_daemon())
565 engine->load_deployment(argv[2]);
567 SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, rank_counts);
568 MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name);
570 /* Clean IO before the run */
574 if (MC_is_active()) {
579 xbt_os_walltimer_stop(global_timer);
580 if (simgrid::config::get_value<bool>("smpi/display-timing")) {
581 double global_time = xbt_os_timer_elapsed(global_timer);
582 XBT_INFO("Simulated time: %g seconds. \n\n"
583 "The simulation took %g seconds (after parsing and platform setup)\n"
584 "%g seconds were actual computation of the application",
585 SIMIX_get_clock(), global_time , smpi_total_benched_time);
587 if (smpi_total_benched_time/global_time>=0.75)
588 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
589 "You may want to use sampling functions or trace replay to reduce this.");
594 return smpi_exit_status;
597 // Called either directly from the user code, or from the code called by smpirun
600 simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::Actor& actor) {
601 if (not actor.is_daemon())
602 actor.extension_set<simgrid::smpi::ActorExt>(new simgrid::smpi::ActorExt(&actor));
604 simgrid::s4u::Host::on_creation.connect(
605 [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); });
606 for (auto const& host : simgrid::s4u::Engine::get_instance()->get_all_hosts())
607 host->extension_set(new simgrid::smpi::Host(host));
609 if (not MC_is_active()) {
610 global_timer = xbt_os_timer_new();
611 xbt_os_walltimer_start(global_timer);
614 smpi_check_options();
619 smpi_bench_destroy();
620 smpi_shared_destroy();
621 smpi_deployment_cleanup_instances();
623 if (simgrid::smpi::colls::smpi_coll_cleanup_callback != nullptr)
624 simgrid::smpi::colls::smpi_coll_cleanup_callback();
626 MPI_COMM_WORLD = MPI_COMM_NULL;
628 if (not MC_is_active()) {
629 xbt_os_timer_free(global_timer);
632 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
633 smpi_destroy_global_memory_segments();
634 if (simgrid::smpi::F2C::lookup() != nullptr)
635 simgrid::smpi::F2C::delete_lookup();
638 void smpi_mpi_init() {
639 smpi_init_fortran_types();
640 if(smpi_init_sleep > 0)
641 simgrid::s4u::this_actor::sleep_for(smpi_init_sleep);
644 void SMPI_thread_create() {
645 TRACE_smpi_init(simgrid::s4u::this_actor::get_pid(), __func__);