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 "src/kernel/activity/CommImpl.hpp"
13 #include "src/simix/smx_private.hpp"
14 #include "src/smpi/include/smpi_actor.hpp"
15 #include "xbt/config.hpp"
18 #include <boost/algorithm/string.hpp> /* split */
19 #include <boost/tokenizer.hpp>
21 #include <cstdint> /* intmax_t */
28 #include "src/mc/mc_config.hpp"
32 #include <sys/sendfile.h>
39 #if not defined(__APPLE__) && not defined(__HAIKU__)
43 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)");
46 extern "C" void for_rtl_init_ (int *, char **);
47 extern "C" void for_rtl_finish_ ();
49 extern "C" void __io_set_argc(int);
50 extern "C" void __io_set_argv(char **);
52 extern "C" void _gfortran_set_args(int, char **);
55 /* RTLD_DEEPBIND is a bad idea of GNU ld that obviously does not exist on other platforms
56 * See https://www.akkadia.org/drepper/dsohowto.pdf
57 * and https://lists.freebsd.org/pipermail/freebsd-current/2016-March/060284.html
59 #if !RTLD_DEEPBIND || HAVE_SANITIZER_ADDRESS || HAVE_SANITIZER_THREAD
60 #define WANT_RTLD_DEEPBIND 0
62 #define WANT_RTLD_DEEPBIND RTLD_DEEPBIND
66 std::string papi_default_config_name = "default";
67 std::map</* computation unit name */ std::string, papi_process_data> units2papi_setup;
70 std::unordered_map<std::string, double> location2speedup;
72 static int smpi_exit_status = 0;
73 extern double smpi_total_benched_time;
74 xbt_os_timer_t global_timer;
75 static std::vector<std::string> privatize_libs_paths;
77 * Setting MPI_COMM_WORLD to MPI_COMM_UNINITIALIZED (it's a variable)
78 * is important because the implementation of MPI_Comm checks
79 * "this == MPI_COMM_UNINITIALIZED"? If yes, it uses smpi_process()->comm_world()
81 * This is basically how we only have one global variable but all processes have
82 * different communicators (the one their SMPI instance uses).
84 * See smpi_comm.cpp and the functions therein for details.
86 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
87 // No instance gets manually created; check also the smpirun.in script as
88 // this default name is used there as well (when the <actor> tag is generated).
89 static const std::string smpi_default_instance_name("smpirun");
90 static simgrid::config::Flag<double> smpi_init_sleep(
91 "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
93 void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*,
94 size_t) = &smpi_comm_copy_buffer_callback;
96 simgrid::smpi::ActorExt* smpi_process()
98 simgrid::s4u::ActorPtr me = simgrid::s4u::Actor::self();
100 if (me == nullptr) // This happens sometimes (eg, when linking against NS3 because it pulls openMPI...)
103 return me->extension<simgrid::smpi::ActorExt>();
106 simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor)
108 if (actor.get() == nullptr)
110 return actor->extension<simgrid::smpi::ActorExt>();
113 MPI_Comm smpi_process_comm_self(){
114 return smpi_process()->comm_self();
117 MPI_Info smpi_process_info_env(){
118 return smpi_process()->info_env();
121 void * smpi_process_get_user_data(){
122 return simgrid::s4u::Actor::self()->get_data();
125 void smpi_process_set_user_data(void *data){
126 simgrid::s4u::Actor::self()->set_data(data);
129 void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
131 static void (*saved_callback)(smx_activity_t, void*, size_t);
132 saved_callback = callback;
133 smpi_comm_copy_data_callback = [](simgrid::kernel::activity::CommImpl* comm, void* buff, size_t size) {
134 saved_callback(smx_activity_t(comm), buff, size);
138 static void memcpy_private(void* dest, const void* src, std::vector<std::pair<size_t, size_t>>& private_blocks)
140 for (auto const& block : private_blocks)
141 memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first);
144 static void check_blocks(std::vector<std::pair<size_t, size_t>> &private_blocks, size_t buff_size) {
145 for (auto const& block : private_blocks)
146 xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
149 static void smpi_cleanup_comm_after_copy(simgrid::kernel::activity::CommImpl* comm, void* buff){
150 if (comm->detached()) {
151 // if this is a detached send, the source buffer was duplicated by SMPI
152 // sender to make the original buffer available to the application ASAP
154 //It seems that the request is used after the call there this should be free somewhere else but where???
155 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
156 comm->src_buff_ = nullptr;
160 void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff, size_t buff_size)
162 size_t src_offset = 0;
163 size_t dst_offset = 0;
164 std::vector<std::pair<size_t, size_t>> src_private_blocks;
165 std::vector<std::pair<size_t, size_t>> dst_private_blocks;
166 XBT_DEBUG("Copy the data over");
167 if(smpi_is_shared(buff, src_private_blocks, &src_offset)) {
168 src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
169 if (src_private_blocks.size()==0){//simple shared malloc ... return.
170 XBT_VERB("Sender is shared. Let's ignore it.");
171 smpi_cleanup_comm_after_copy(comm, buff);
176 src_private_blocks.clear();
177 src_private_blocks.push_back(std::make_pair(0, buff_size));
179 if (smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset)) {
180 dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
181 if (dst_private_blocks.size()==0){//simple shared malloc ... return.
182 XBT_VERB("Receiver is shared. Let's ignore it.");
183 smpi_cleanup_comm_after_copy(comm, buff);
188 dst_private_blocks.clear();
189 dst_private_blocks.push_back(std::make_pair(0, buff_size));
191 check_blocks(src_private_blocks, buff_size);
192 check_blocks(dst_private_blocks, buff_size);
193 auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
194 check_blocks(private_blocks, buff_size);
196 if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) &&
197 (static_cast<char*>(buff) >= smpi_data_exe_start) &&
198 (static_cast<char*>(buff) < smpi_data_exe_start + smpi_data_exe_size)) {
199 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
200 smpi_switch_data_segment(comm->src_actor_->iface());
201 tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
202 memcpy_private(tmpbuff, buff, private_blocks);
205 if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) &&
206 ((char*)comm->dst_buff_ >= smpi_data_exe_start) &&
207 ((char*)comm->dst_buff_ < smpi_data_exe_start + smpi_data_exe_size)) {
208 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
209 smpi_switch_data_segment(comm->dst_actor_->iface());
211 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_);
212 memcpy_private(comm->dst_buff_, tmpbuff, private_blocks);
214 smpi_cleanup_comm_after_copy(comm,buff);
219 void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, void*, size_t)
221 /* nothing done in this version */
225 return MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED;
228 static void smpi_init_papi()
231 // This map holds for each computation unit (such as "default" or "process1" etc.)
232 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
233 // and the (computed) event_set.
235 if (not smpi_cfg_papi_events_file().empty()) {
236 if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
237 XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?"
238 " Expected version is %u", PAPI_VER_CURRENT);
240 typedef boost::tokenizer<boost::char_separator<char>> Tokenizer;
241 boost::char_separator<char> separator_units(";");
242 std::string str = smpi_cfg_papi_events_file();
243 Tokenizer tokens(str, separator_units);
245 // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it.
246 // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units".
247 for (auto const& 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 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
275 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
276 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
279 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
281 counters2values.push_back(
282 // We cannot just pass *events_it, as this is of type const basic_string
283 std::make_pair<std::string, long long>(std::string(*events_it), 0));
286 std::string unit_name = *(event_tokens.begin());
287 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
289 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
297 typedef std::function<int(int argc, char *argv[])> smpi_entry_point_type;
298 typedef int (* smpi_c_entry_point_type)(int argc, char **argv);
299 typedef void (*smpi_fortran_entry_point_type)();
301 template <typename F>
302 static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path, std::vector<std::string> args)
304 // copy C strings, we need them writable
305 std::vector<char*>* args4argv = new std::vector<char*>(args.size());
306 std::transform(begin(args), end(args), begin(*args4argv), [](const std::string& s) { return xbt_strdup(s.c_str()); });
308 // set argv[0] to executable_path
309 xbt_free((*args4argv)[0]);
310 (*args4argv)[0] = xbt_strdup(executable_path.c_str());
313 // take a copy of args4argv to keep reference of the allocated strings
314 const std::vector<char*> args2str(*args4argv);
316 int argc = args4argv->size();
317 args4argv->push_back(nullptr);
318 char** argv = args4argv->data();
321 for_rtl_init_ (&argc, argv);
326 _gfortran_set_args(argc, argv);
328 int res = entry_point(argc, argv);
333 for (char* s : args2str)
339 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
340 if (smpi_exit_status == 0)
341 smpi_exit_status = res;
347 // TODO, remove the number of functions involved here
348 static smpi_entry_point_type smpi_resolve_function(void* handle)
350 smpi_fortran_entry_point_type entry_point_fortran = (smpi_fortran_entry_point_type)dlsym(handle, "user_main_");
351 if (entry_point_fortran != nullptr) {
352 return [entry_point_fortran](int, char**) {
353 entry_point_fortran();
358 smpi_c_entry_point_type entry_point = (smpi_c_entry_point_type)dlsym(handle, "main");
359 if (entry_point != nullptr) {
363 return smpi_entry_point_type();
366 static void smpi_copy_file(const std::string& src, const std::string& target, off_t fdin_size)
368 int fdin = open(src.c_str(), O_RDONLY);
369 xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str());
370 int fdout = open(target.c_str(), O_CREAT | O_RDWR, S_IRWXU);
371 xbt_assert(fdout >= 0, "Cannot write into %s", target.c_str());
373 XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast<intmax_t>(fdin_size), target.c_str());
375 ssize_t sent_size = sendfile(fdout, fdin, NULL, fdin_size);
376 if (sent_size == fdin_size) {
380 } else if (sent_size != -1 || errno != ENOSYS) {
381 xbt_die("Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(),
382 sent_size, static_cast<intmax_t>(fdin_size), errno, strerror(errno));
385 // If this point is reached, sendfile() actually is not available. Copy file by hand.
386 const int bufsize = 1024 * 1024 * 4;
387 char* buf = new char[bufsize];
388 while (int got = read(fdin, buf, bufsize)) {
390 xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str());
394 while (int done = write(fdout, p, todo)) {
396 xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str());
409 #if not defined(__APPLE__) && not defined(__HAIKU__)
410 static int visit_libs(struct dl_phdr_info* info, size_t, void* data)
412 char* libname = (char*)(data);
413 const char *path = info->dlpi_name;
414 if(strstr(path, libname)){
415 strncpy(libname, path, 512);
423 static void smpi_init_privatization_dlopen(const std::string& executable)
425 // Prepare the copy of the binary (get its size)
426 struct stat fdin_stat;
427 stat(executable.c_str(), &fdin_stat);
428 off_t fdin_size = fdin_stat.st_size;
430 std::string libnames = simgrid::config::get_value<std::string>("smpi/privatize-libs");
431 if (not libnames.empty()) {
433 std::vector<std::string> privatize_libs;
434 boost::split(privatize_libs, libnames, boost::is_any_of(";"));
436 for (auto const& libname : privatize_libs) {
437 // load the library once to add it to the local libs, to get the absolute path
438 void* libhandle = dlopen(libname.c_str(), RTLD_LAZY);
439 // get library name from path
440 char fullpath[512] = {'\0'};
441 strncpy(fullpath, libname.c_str(), 511);
442 #if not defined(__APPLE__) && not defined(__HAIKU__)
443 xbt_assert(0 != dl_iterate_phdr(visit_libs, fullpath),
444 "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath);
445 XBT_DEBUG("Extra lib to privatize '%s' found", fullpath);
447 xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku");
449 privatize_libs_paths.push_back(fullpath);
454 simix_global->default_function = [executable, fdin_size](std::vector<std::string> args) {
455 return std::function<void()>([executable, fdin_size, args] {
456 static std::size_t rank = 0;
457 // Copy the dynamic library:
458 std::string target_executable =
459 executable + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so";
461 smpi_copy_file(executable, target_executable, fdin_size);
462 // if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one.
463 std::vector<std::string> target_libs;
464 for (auto const& libpath : privatize_libs_paths) {
465 // if we were given a full path, strip it
466 size_t index = libpath.find_last_of("/\\");
468 if (index != std::string::npos)
469 libname = libpath.substr(index + 1);
471 if (not libname.empty()) {
472 // load the library to add it to the local libs, to get the absolute path
473 struct stat fdin_stat2;
474 stat(libpath.c_str(), &fdin_stat2);
475 off_t fdin_size2 = fdin_stat2.st_size;
477 // Copy the dynamic library, the new name must be the same length as the old one
478 // just replace the name with 7 digits for the rank and the rest of the name.
479 unsigned int pad = 7;
480 if (libname.length() < pad)
481 pad = libname.length();
482 std::string target_lib =
483 std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
484 target_libs.push_back(target_lib);
485 XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2);
486 smpi_copy_file(libpath, target_lib, fdin_size2);
488 std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_lib + "/g' " + target_executable;
489 xbt_assert(system(sedcommand.c_str()) == 0, "error while applying sed command %s \n", sedcommand.c_str());
494 // Load the copy and resolve the entry point:
495 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND);
496 int saved_errno = errno;
497 if (simgrid::config::get_value<bool>("smpi/keep-temps") == false) {
498 unlink(target_executable.c_str());
499 for (const std::string& target_lib : target_libs)
500 unlink(target_lib.c_str());
502 xbt_assert(handle != nullptr, "dlopen failed: %s (errno: %d -- %s)", dlerror(), saved_errno,
503 strerror(saved_errno));
505 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
506 xbt_assert(entry_point, "Could not resolve entry point");
507 smpi_run_entry_point(entry_point, executable, args);
512 static void smpi_init_privatization_no_dlopen(const std::string& executable)
514 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
515 smpi_prepare_global_memory_segment();
517 // Load the dynamic library and resolve the entry point:
518 void* handle = dlopen(executable.c_str(), RTLD_LAZY | RTLD_LOCAL);
519 xbt_assert(handle != nullptr, "dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno,
521 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
522 xbt_assert(entry_point, "main not found in %s", executable.c_str());
524 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
525 smpi_backup_global_memory_segment();
527 // Execute the same entry point for each simulated process:
528 simix_global->default_function = [entry_point, executable](std::vector<std::string> args) {
529 return std::function<void()>(
530 [entry_point, executable, args] { smpi_run_entry_point(entry_point, executable, args); });
534 int smpi_main(const char* executable, int argc, char* argv[])
536 if (getenv("SMPI_PRETEND_CC") != nullptr) {
537 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
538 * configuration tools */
542 SMPI_switch_data_segment = &smpi_switch_data_segment;
545 SIMIX_global_init(&argc, argv);
547 auto engine = simgrid::s4u::Engine::get_instance();
549 sg_storage_file_system_init();
550 // parse the platform file: get the host list
551 engine->load_platform(argv[1]);
552 SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback);
554 if (smpi_cfg_privatization() == SmpiPrivStrategies::DLOPEN)
555 smpi_init_privatization_dlopen(executable);
557 smpi_init_privatization_no_dlopen(executable);
559 simgrid::smpi::colls::set_collectives();
560 simgrid::smpi::colls::smpi_coll_cleanup_callback = nullptr;
564 /* This is a ... heavy way to count the MPI ranks */
566 simgrid::s4u::Actor::on_creation.connect([&rank_counts](simgrid::s4u::Actor& actor) {
567 if (not actor.is_daemon())
570 engine->load_deployment(argv[2]);
572 SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, rank_counts);
573 MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name);
575 /* Clean IO before the run */
579 if (MC_is_active()) {
585 xbt_os_walltimer_stop(global_timer);
586 if (simgrid::config::get_value<bool>("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.");
600 return smpi_exit_status;
603 // Called either directly from the user code, or from the code called by smpirun
606 simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::Actor& actor) {
607 if (not actor.is_daemon())
608 actor.extension_set<simgrid::smpi::ActorExt>(new simgrid::smpi::ActorExt(&actor));
610 simgrid::s4u::Host::on_creation.connect(
611 [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); });
612 for (auto const& host : simgrid::s4u::Engine::get_instance()->get_all_hosts())
613 host->extension_set(new simgrid::smpi::Host(host));
615 if (not MC_is_active()) {
616 global_timer = xbt_os_timer_new();
617 xbt_os_walltimer_start(global_timer);
620 smpi_check_options();
625 smpi_bench_destroy();
626 smpi_shared_destroy();
627 smpi_deployment_cleanup_instances();
629 if (simgrid::smpi::colls::smpi_coll_cleanup_callback != nullptr)
630 simgrid::smpi::colls::smpi_coll_cleanup_callback();
632 MPI_COMM_WORLD = MPI_COMM_NULL;
634 if (not MC_is_active()) {
635 xbt_os_timer_free(global_timer);
638 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
639 smpi_destroy_global_memory_segments();
640 if (simgrid::smpi::F2C::lookup() != nullptr)
641 simgrid::smpi::F2C::delete_lookup();
644 void smpi_mpi_init() {
645 smpi_init_fortran_types();
646 if(smpi_init_sleep > 0)
647 simgrid::s4u::this_actor::sleep_for(smpi_init_sleep);