1 /* Copyright (c) 2007-2019. 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(smx_activity_t(comm), buff, size);
135 static void memcpy_private(void* dest, const void* src, 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(std::vector<std::pair<size_t, size_t>> &private_blocks, size_t buff_size) {
142 for (auto const& block : private_blocks)
143 xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
146 static void smpi_cleanup_comm_after_copy(simgrid::kernel::activity::CommImpl* comm, void* buff){
147 if (comm->detached()) {
148 // if this is a detached send, the source buffer was duplicated by SMPI
149 // sender to make the original buffer available to the application ASAP
151 //It seems that the request is used after the call there this should be free somewhere else but where???
152 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
153 comm->src_buff_ = nullptr;
157 void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff, size_t buff_size)
159 size_t src_offset = 0;
160 size_t dst_offset = 0;
161 std::vector<std::pair<size_t, size_t>> src_private_blocks;
162 std::vector<std::pair<size_t, size_t>> dst_private_blocks;
163 XBT_DEBUG("Copy the data over");
164 if(smpi_is_shared(buff, src_private_blocks, &src_offset)) {
165 src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
166 if (src_private_blocks.size()==0){//simple shared malloc ... return.
167 XBT_VERB("Sender is shared. Let's ignore it.");
168 smpi_cleanup_comm_after_copy(comm, buff);
173 src_private_blocks.clear();
174 src_private_blocks.push_back(std::make_pair(0, buff_size));
176 if (smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset)) {
177 dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
178 if (dst_private_blocks.size()==0){//simple shared malloc ... return.
179 XBT_VERB("Receiver is shared. Let's ignore it.");
180 smpi_cleanup_comm_after_copy(comm, buff);
185 dst_private_blocks.clear();
186 dst_private_blocks.push_back(std::make_pair(0, buff_size));
188 check_blocks(src_private_blocks, buff_size);
189 check_blocks(dst_private_blocks, buff_size);
190 auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
191 check_blocks(private_blocks, buff_size);
193 if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) &&
194 (static_cast<char*>(buff) >= smpi_data_exe_start) &&
195 (static_cast<char*>(buff) < smpi_data_exe_start + smpi_data_exe_size)) {
196 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
197 smpi_switch_data_segment(comm->src_actor_->iface());
198 tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
199 memcpy_private(tmpbuff, buff, private_blocks);
202 if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) &&
203 ((char*)comm->dst_buff_ >= smpi_data_exe_start) &&
204 ((char*)comm->dst_buff_ < smpi_data_exe_start + smpi_data_exe_size)) {
205 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
206 smpi_switch_data_segment(comm->dst_actor_->iface());
208 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_);
209 memcpy_private(comm->dst_buff_, tmpbuff, private_blocks);
211 smpi_cleanup_comm_after_copy(comm,buff);
216 void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, void*, size_t)
218 /* nothing done in this version */
222 return MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED;
225 static void smpi_init_papi()
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.
232 if (not smpi_cfg_papi_events_file().empty()) {
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 %u", PAPI_VER_CURRENT);
237 typedef boost::tokenizer<boost::char_separator<char>> Tokenizer;
238 boost::char_separator<char> separator_units(";");
239 std::string str = smpi_cfg_papi_events_file();
240 Tokenizer tokens(str, separator_units);
242 // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it.
243 // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units".
244 for (auto const& unit_it : tokens) {
245 boost::char_separator<char> separator_events(":");
246 Tokenizer event_tokens(unit_it, separator_events);
248 int event_set = PAPI_NULL;
249 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
250 // TODO: Should this let the whole simulation die?
251 XBT_CRITICAL("Could not create PAPI event set during init.");
254 // NOTE: We cannot use a map here, as we must obey the order of the counters
255 // This is important for PAPI: We need to map the values of counters back
256 // to the event_names (so, when PAPI_read() has finished)!
257 papi_counter_t counters2values;
259 // Iterate over all counters that were specified for this specific
261 // Note that we need to remove the name of the unit
262 // (that could also be the "default" value), which always comes first.
263 // Hence, we start at ++(events.begin())!
264 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) {
266 int event_code = PAPI_NULL;
267 char* event_name = const_cast<char*>((*events_it).c_str());
268 if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) {
269 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
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);
278 counters2values.push_back(
279 // We cannot just pass *events_it, as this is of type const basic_string
280 std::make_pair(std::string(*events_it), 0LL));
283 std::string unit_name = *(event_tokens.begin());
284 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
286 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
294 typedef std::function<int(int argc, char *argv[])> smpi_entry_point_type;
295 typedef int (* smpi_c_entry_point_type)(int argc, char **argv);
296 typedef void (*smpi_fortran_entry_point_type)();
298 template <typename F>
299 static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path, std::vector<std::string> args)
301 // copy C strings, we need them writable
302 std::vector<char*>* args4argv = new std::vector<char*>(args.size());
303 std::transform(begin(args), end(args), begin(*args4argv), [](const std::string& s) { return xbt_strdup(s.c_str()); });
305 // set argv[0] to executable_path
306 xbt_free((*args4argv)[0]);
307 (*args4argv)[0] = xbt_strdup(executable_path.c_str());
310 // take a copy of args4argv to keep reference of the allocated strings
311 const std::vector<char*> args2str(*args4argv);
313 int argc = args4argv->size();
314 args4argv->push_back(nullptr);
315 char** argv = args4argv->data();
318 for_rtl_init_ (&argc, argv);
323 _gfortran_set_args(argc, argv);
325 int res = entry_point(argc, argv);
330 for (char* s : args2str)
336 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
337 if (smpi_exit_status == 0)
338 smpi_exit_status = res;
344 // TODO, remove the number of functions involved here
345 static smpi_entry_point_type smpi_resolve_function(void* handle)
347 smpi_fortran_entry_point_type entry_point_fortran = (smpi_fortran_entry_point_type)dlsym(handle, "user_main_");
348 if (entry_point_fortran != nullptr) {
349 return [entry_point_fortran](int, char**) {
350 entry_point_fortran();
355 smpi_c_entry_point_type entry_point = (smpi_c_entry_point_type)dlsym(handle, "main");
356 if (entry_point != nullptr) {
360 return smpi_entry_point_type();
363 static void smpi_copy_file(const std::string& src, const std::string& target, off_t fdin_size)
365 int fdin = open(src.c_str(), O_RDONLY);
366 xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str());
367 int fdout = open(target.c_str(), O_CREAT | O_RDWR, S_IRWXU);
368 xbt_assert(fdout >= 0, "Cannot write into %s", target.c_str());
370 XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast<intmax_t>(fdin_size), target.c_str());
372 ssize_t sent_size = sendfile(fdout, fdin, NULL, fdin_size);
373 if (sent_size == fdin_size) {
377 } else if (sent_size != -1 || errno != ENOSYS) {
378 xbt_die("Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(),
379 sent_size, static_cast<intmax_t>(fdin_size), errno, strerror(errno));
382 // If this point is reached, sendfile() actually is not available. Copy file by hand.
383 const int bufsize = 1024 * 1024 * 4;
384 char* buf = new char[bufsize];
385 while (int got = read(fdin, buf, bufsize)) {
387 xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str());
391 while (int done = write(fdout, p, todo)) {
393 xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str());
406 #if not defined(__APPLE__) && not defined(__HAIKU__)
407 static int visit_libs(struct dl_phdr_info* info, size_t, void* data)
409 char* libname = (char*)(data);
410 const char *path = info->dlpi_name;
411 if(strstr(path, libname)){
412 strncpy(libname, path, 512);
420 static void smpi_init_privatization_dlopen(const std::string& executable)
422 // Prepare the copy of the binary (get its size)
423 struct stat fdin_stat;
424 stat(executable.c_str(), &fdin_stat);
425 off_t fdin_size = fdin_stat.st_size;
427 std::string libnames = simgrid::config::get_value<std::string>("smpi/privatize-libs");
428 if (not libnames.empty()) {
430 std::vector<std::string> privatize_libs;
431 boost::split(privatize_libs, libnames, boost::is_any_of(";"));
433 for (auto const& libname : privatize_libs) {
434 // load the library once to add it to the local libs, to get the absolute path
435 void* libhandle = dlopen(libname.c_str(), RTLD_LAZY);
436 // get library name from path
437 char fullpath[512] = {'\0'};
438 strncpy(fullpath, libname.c_str(), 511);
439 #if not defined(__APPLE__) && not defined(__HAIKU__)
440 xbt_assert(0 != dl_iterate_phdr(visit_libs, fullpath),
441 "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath);
442 XBT_DEBUG("Extra lib to privatize '%s' found", fullpath);
444 xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku");
446 privatize_libs_paths.push_back(fullpath);
451 simix_global->default_function = [executable, fdin_size](std::vector<std::string> args) {
452 return std::function<void()>([executable, fdin_size, args] {
453 static std::size_t rank = 0;
454 // Copy the dynamic library:
455 std::string target_executable =
456 executable + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so";
458 smpi_copy_file(executable, target_executable, fdin_size);
459 // if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one.
460 std::vector<std::string> target_libs;
461 for (auto const& libpath : privatize_libs_paths) {
462 // if we were given a full path, strip it
463 size_t index = libpath.find_last_of("/\\");
465 if (index != std::string::npos)
466 libname = libpath.substr(index + 1);
468 if (not libname.empty()) {
469 // load the library to add it to the local libs, to get the absolute path
470 struct stat fdin_stat2;
471 stat(libpath.c_str(), &fdin_stat2);
472 off_t fdin_size2 = fdin_stat2.st_size;
474 // Copy the dynamic library, the new name must be the same length as the old one
475 // just replace the name with 7 digits for the rank and the rest of the name.
476 unsigned int pad = 7;
477 if (libname.length() < pad)
478 pad = libname.length();
479 std::string target_lib =
480 std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
481 target_libs.push_back(target_lib);
482 XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2);
483 smpi_copy_file(libpath, target_lib, fdin_size2);
485 std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_lib + "/g' " + target_executable;
486 xbt_assert(system(sedcommand.c_str()) == 0, "error while applying sed command %s \n", sedcommand.c_str());
491 // Load the copy and resolve the entry point:
492 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND);
493 int saved_errno = errno;
494 if (simgrid::config::get_value<bool>("smpi/keep-temps") == false) {
495 unlink(target_executable.c_str());
496 for (const std::string& target_lib : target_libs)
497 unlink(target_lib.c_str());
499 xbt_assert(handle != nullptr, "dlopen failed: %s (errno: %d -- %s)", dlerror(), saved_errno,
500 strerror(saved_errno));
502 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
503 xbt_assert(entry_point, "Could not resolve entry point");
504 smpi_run_entry_point(entry_point, executable, args);
509 static void smpi_init_privatization_no_dlopen(const std::string& executable)
511 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
512 smpi_prepare_global_memory_segment();
514 // Load the dynamic library and resolve the entry point:
515 void* handle = dlopen(executable.c_str(), RTLD_LAZY | RTLD_LOCAL);
516 xbt_assert(handle != nullptr, "dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno,
518 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
519 xbt_assert(entry_point, "main not found in %s", executable.c_str());
521 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
522 smpi_backup_global_memory_segment();
524 // Execute the same entry point for each simulated process:
525 simix_global->default_function = [entry_point, executable](std::vector<std::string> args) {
526 return std::function<void()>(
527 [entry_point, executable, args] { smpi_run_entry_point(entry_point, executable, args); });
531 int smpi_main(const char* executable, int argc, char* argv[])
533 if (getenv("SMPI_PRETEND_CC") != nullptr) {
534 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
535 * configuration tools */
539 SMPI_switch_data_segment = &smpi_switch_data_segment;
542 SIMIX_global_init(&argc, argv);
544 auto engine = simgrid::s4u::Engine::get_instance();
546 sg_storage_file_system_init();
547 // parse the platform file: get the host list
548 engine->load_platform(argv[1]);
549 SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback);
551 if (smpi_cfg_privatization() == SmpiPrivStrategies::DLOPEN)
552 smpi_init_privatization_dlopen(executable);
554 smpi_init_privatization_no_dlopen(executable);
556 simgrid::smpi::colls::set_collectives();
557 simgrid::smpi::colls::smpi_coll_cleanup_callback = nullptr;
561 /* This is a ... heavy way to count the MPI ranks */
563 simgrid::s4u::Actor::on_creation.connect([&rank_counts](simgrid::s4u::Actor& actor) {
564 if (not actor.is_daemon())
567 engine->load_deployment(argv[2]);
569 SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, rank_counts);
570 MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name);
572 /* Clean IO before the run */
576 if (MC_is_active()) {
582 xbt_os_walltimer_stop(global_timer);
583 if (simgrid::config::get_value<bool>("smpi/display-timing")) {
584 double global_time = xbt_os_timer_elapsed(global_timer);
585 XBT_INFO("Simulated time: %g seconds. \n\n"
586 "The simulation took %g seconds (after parsing and platform setup)\n"
587 "%g seconds were actual computation of the application",
588 SIMIX_get_clock(), global_time , smpi_total_benched_time);
590 if (smpi_total_benched_time/global_time>=0.75)
591 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
592 "You may want to use sampling functions or trace replay to reduce this.");
597 return smpi_exit_status;
600 // Called either directly from the user code, or from the code called by smpirun
603 simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::Actor& actor) {
604 if (not actor.is_daemon())
605 actor.extension_set<simgrid::smpi::ActorExt>(new simgrid::smpi::ActorExt(&actor));
607 simgrid::s4u::Host::on_creation.connect(
608 [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); });
609 for (auto const& host : simgrid::s4u::Engine::get_instance()->get_all_hosts())
610 host->extension_set(new simgrid::smpi::Host(host));
612 if (not MC_is_active()) {
613 global_timer = xbt_os_timer_new();
614 xbt_os_walltimer_start(global_timer);
617 smpi_check_options();
622 smpi_bench_destroy();
623 smpi_shared_destroy();
624 smpi_deployment_cleanup_instances();
626 if (simgrid::smpi::colls::smpi_coll_cleanup_callback != nullptr)
627 simgrid::smpi::colls::smpi_coll_cleanup_callback();
629 MPI_COMM_WORLD = MPI_COMM_NULL;
631 if (not MC_is_active()) {
632 xbt_os_timer_free(global_timer);
635 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
636 smpi_destroy_global_memory_segments();
637 if (simgrid::smpi::F2C::lookup() != nullptr)
638 simgrid::smpi::F2C::delete_lookup();
641 void smpi_mpi_init() {
642 smpi_init_fortran_types();
643 if(smpi_init_sleep > 0)
644 simgrid::s4u::this_actor::sleep_for(smpi_init_sleep);