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. */
9 #include "simgrid/s4u/Mailbox.hpp"
10 #include "smpi/smpi_shared_malloc.hpp"
11 #include "simgrid/sg_config.h"
12 #include "src/kernel/activity/SynchroComm.hpp"
13 #include "src/mc/mc_record.h"
14 #include "src/mc/mc_replay.h"
15 #include "src/msg/msg_private.h"
16 #include "src/simix/smx_private.h"
17 #include "surf/surf.h"
18 #include "xbt/replay.hpp"
19 #include <xbt/config.hpp>
21 #include <float.h> /* DBL_MAX */
30 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)");
31 #include <boost/tokenizer.hpp>
32 #include <boost/algorithm/string.hpp> /* trim_right / trim_left */
36 const char* papi_default_config_name = "default";
38 struct papi_process_data {
39 papi_counter_t counter_data;
44 std::unordered_map<std::string, double> location2speedup;
46 simgrid::smpi::Process **process_data = nullptr;
47 int process_count = 0;
48 int smpi_universe_size = 0;
49 int* index_to_process_data = nullptr;
50 extern double smpi_total_benched_time;
51 xbt_os_timer_t global_timer;
52 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
53 MPI_Errhandler *MPI_ERRORS_RETURN = nullptr;
54 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = nullptr;
55 MPI_Errhandler *MPI_ERRHANDLER_NULL = nullptr;
56 static simgrid::config::Flag<double> smpi_wtime_sleep(
57 "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0);
58 static simgrid::config::Flag<double> smpi_init_sleep(
59 "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
61 void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t) = &smpi_comm_copy_buffer_callback;
65 int smpi_process_count()
70 simgrid::smpi::Process* smpi_process()
72 simgrid::MsgActorExt* msgExt = static_cast<simgrid::MsgActorExt*>(SIMIX_process_self()->data);
73 return static_cast<simgrid::smpi::Process*>(msgExt->data);
76 simgrid::smpi::Process* smpi_process_remote(int index)
78 return process_data[index_to_process_data[index]];
81 MPI_Comm smpi_process_comm_self(){
82 return smpi_process()->comm_self();
85 void smpi_process_init(int *argc, char ***argv){
86 simgrid::smpi::Process::init(argc, argv);
89 int smpi_process_index(){
90 return smpi_process()->index();
94 int smpi_global_size()
96 char *value = getenv("SMPI_GLOBAL_SIZE");
97 xbt_assert(value,"Please set env var SMPI_GLOBAL_SIZE to the expected number of processes.");
99 return xbt_str_parse_int(value, "SMPI_GLOBAL_SIZE contains a non-numerical value: %s");
102 void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
104 smpi_comm_copy_data_callback = callback;
107 static void print(std::vector<std::pair<size_t, size_t>> vec) {
108 fprintf(stderr, "{");
110 fprintf(stderr, "(0x%lx, 0x%lx),", elt.first, elt.second);
112 fprintf(stderr, "}\n");
114 static void memcpy_private(void *dest, const void *src, size_t n, std::vector<std::pair<size_t, size_t>> &private_blocks) {
115 for(auto block : private_blocks) {
116 memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first);
120 static void check_blocks(std::vector<std::pair<size_t, size_t>> &private_blocks, size_t buff_size) {
121 for(auto block : private_blocks) {
122 xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
126 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);
129 int src_shared=0, dst_shared=0;
130 size_t src_offset=0, dst_offset=0;
131 std::vector<std::pair<size_t, size_t>> src_private_blocks;
132 std::vector<std::pair<size_t, size_t>> dst_private_blocks;
133 XBT_DEBUG("Copy the data over");
134 if((src_shared=smpi_is_shared(buff, src_private_blocks, &src_offset))) {
135 XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff);
136 src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
139 src_private_blocks.clear();
140 src_private_blocks.push_back(std::make_pair(0, buff_size));
142 if((dst_shared=smpi_is_shared((char*)comm->dst_buff, dst_private_blocks, &dst_offset))) {
143 XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff);
144 dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
147 dst_private_blocks.clear();
148 dst_private_blocks.push_back(std::make_pair(0, buff_size));
151 fprintf(stderr, "size: 0x%x\n", buff_size);
152 fprintf(stderr, "src: ");
153 print(src_private_blocks);
154 fprintf(stderr, "src_offset = 0x%x\n", src_offset);
155 fprintf(stderr, "dst: ");
156 print(dst_private_blocks);
157 fprintf(stderr, "dst_offset = 0x%x\n", dst_offset);
159 check_blocks(src_private_blocks, buff_size);
160 check_blocks(dst_private_blocks, buff_size);
161 auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
163 fprintf(stderr, "Private blocks: ");
164 print(private_blocks);
166 check_blocks(private_blocks, buff_size);
168 if((smpi_privatize_global_variables) && (static_cast<char*>(buff) >= smpi_start_data_exe)
169 && (static_cast<char*>(buff) < smpi_start_data_exe + smpi_size_data_exe )
171 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
173 smpi_switch_data_segment(
174 (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->src_proc->data)->data))->index()));
175 tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
176 memcpy_private(tmpbuff, buff, buff_size, private_blocks);
179 if((smpi_privatize_global_variables) && ((char*)comm->dst_buff >= smpi_start_data_exe)
180 && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){
181 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
182 smpi_switch_data_segment(
183 (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->dst_proc->data)->data))->index()));
186 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
187 memcpy_private(comm->dst_buff, tmpbuff, buff_size, private_blocks);
189 if (comm->detached) {
190 // if this is a detached send, the source buffer was duplicated by SMPI
191 // sender to make the original buffer available to the application ASAP
193 //It seems that the request is used after the call there this should be free somewhere else but where???
194 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
195 comm->src_buff = nullptr;
197 if(tmpbuff!=buff)xbt_free(tmpbuff);
201 void smpi_comm_null_copy_buffer_callback(smx_activity_t comm, void *buff, size_t buff_size)
203 /* nothing done in this version */
206 static void smpi_check_options(){
207 //check correctness of MPI parameters
209 xbt_assert(xbt_cfg_get_int("smpi/async-small-thresh") <= xbt_cfg_get_int("smpi/send-is-detached-thresh"));
211 if (xbt_cfg_is_default_value("smpi/host-speed")) {
212 XBT_INFO("You did not set the power of the host running the simulation. "
213 "The timings will certainly not be accurate. "
214 "Use the option \"--cfg=smpi/host-speed:<flops>\" to set its value."
215 "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
218 xbt_assert(xbt_cfg_get_double("smpi/cpu-threshold") >=0,
219 "The 'smpi/cpu-threshold' option cannot have negative values [anymore]. If you want to discard "
220 "the simulation of any computation, please use 'smpi/simulate-computation:no' instead.");
224 return process_data != nullptr;
227 void smpi_global_init()
233 if (!MC_is_active()) {
234 global_timer = xbt_os_timer_new();
235 xbt_os_walltimer_start(global_timer);
238 if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') {
239 std::string filename {xbt_cfg_get_string("smpi/comp-adjustment-file")};
240 std::ifstream fstream(filename);
241 if (!fstream.is_open()) {
242 xbt_die("Could not open file %s. Does it exist?", filename.c_str());
246 typedef boost::tokenizer< boost::escaped_list_separator<char>> Tokenizer;
247 std::getline(fstream, line); // Skip the header line
248 while (std::getline(fstream, line)) {
250 Tokenizer::iterator it = tok.begin();
251 Tokenizer::iterator end = std::next(tok.begin());
253 std::string location = *it;
254 boost::trim(location);
255 location2speedup.insert(std::pair<std::string, double>(location, std::stod(*end)));
260 // This map holds for each computation unit (such as "default" or "process1" etc.)
261 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
262 // and the (computed) event_set.
263 std::map</* computation unit name */ std::string, papi_process_data> units2papi_setup;
265 if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') {
266 if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
267 XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?"
268 " Expected version is %i",
271 typedef boost::tokenizer<boost::char_separator<char>> Tokenizer;
272 boost::char_separator<char> separator_units(";");
273 std::string str = std::string(xbt_cfg_get_string("smpi/papi-events"));
274 Tokenizer tokens(str, separator_units);
276 // Iterate over all the computational units. This could be
277 // processes, hosts, threads, ranks... You name it. I'm not exactly
278 // sure what we will support eventually, so I'll leave it at the
279 // general term "units".
280 for (auto& unit_it : tokens) {
281 boost::char_separator<char> separator_events(":");
282 Tokenizer event_tokens(unit_it, separator_events);
284 int event_set = PAPI_NULL;
285 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
286 // TODO: Should this let the whole simulation die?
287 XBT_CRITICAL("Could not create PAPI event set during init.");
290 // NOTE: We cannot use a map here, as we must obey the order of the counters
291 // This is important for PAPI: We need to map the values of counters back
292 // to the event_names (so, when PAPI_read() has finished)!
293 papi_counter_t counters2values;
295 // Iterate over all counters that were specified for this specific
297 // Note that we need to remove the name of the unit
298 // (that could also be the "default" value), which always comes first.
299 // Hence, we start at ++(events.begin())!
300 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); events_it++) {
302 int event_code = PAPI_NULL;
303 char* event_name = const_cast<char*>((*events_it).c_str());
304 if (PAPI_event_name_to_code(event_name, &event_code) == PAPI_OK) {
305 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
306 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
309 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
312 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
316 counters2values.push_back(
317 // We cannot just pass *events_it, as this is of type const basic_string
318 std::make_pair<std::string, long long>(std::string(*events_it), 0));
321 std::string unit_name = *(event_tokens.begin());
322 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
324 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
328 if (process_count == 0){
329 process_count = SIMIX_process_count();
332 smpi_universe_size = process_count;
333 process_data = new simgrid::smpi::Process*[process_count];
334 for (i = 0; i < process_count; i++) {
335 process_data[i] = new simgrid::smpi::Process(i);
337 //if the process was launched through smpirun script we generate a global mpi_comm_world
338 //if not, we let MPI_COMM_NULL, and the comm world will be private to each mpi instance
340 group = new simgrid::smpi::Group(process_count);
341 MPI_COMM_WORLD = new simgrid::smpi::Comm(group, nullptr);
342 MPI_Attr_put(MPI_COMM_WORLD, MPI_UNIVERSE_SIZE, reinterpret_cast<void *>(process_count));
343 msg_bar_t bar = MSG_barrier_init(process_count);
345 for (i = 0; i < process_count; i++) {
346 group->set_mapping(i, i);
347 process_data[i]->set_finalization_barrier(bar);
352 void smpi_global_destroy()
354 int count = smpi_process_count();
356 smpi_bench_destroy();
357 smpi_shared_destroy();
358 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
359 delete MPI_COMM_WORLD->group();
360 MSG_barrier_destroy(process_data[0]->finalization_barrier());
362 smpi_deployment_cleanup_instances();
364 for (int i = 0; i < count; i++) {
365 if(process_data[i]->comm_self()!=MPI_COMM_NULL){
366 simgrid::smpi::Comm::destroy(process_data[i]->comm_self());
368 if(process_data[i]->comm_intra()!=MPI_COMM_NULL){
369 simgrid::smpi::Comm::destroy(process_data[i]->comm_intra());
371 xbt_os_timer_free(process_data[i]->timer());
372 xbt_mutex_destroy(process_data[i]->mailboxes_mutex());
373 delete process_data[i];
375 delete[] process_data;
376 process_data = nullptr;
378 if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
379 MPI_COMM_WORLD->cleanup_smp();
380 MPI_COMM_WORLD->cleanup_attr<simgrid::smpi::Comm>();
381 if(simgrid::smpi::Colls::smpi_coll_cleanup_callback!=nullptr)
382 simgrid::smpi::Colls::smpi_coll_cleanup_callback();
383 delete MPI_COMM_WORLD;
386 MPI_COMM_WORLD = MPI_COMM_NULL;
388 if (!MC_is_active()) {
389 xbt_os_timer_free(global_timer);
392 xbt_free(index_to_process_data);
393 if(smpi_privatize_global_variables)
394 smpi_destroy_global_memory_segments();
402 void __attribute__ ((weak)) user_main_()
404 xbt_die("Should not be in this smpi_simulated_main");
407 int __attribute__ ((weak)) smpi_simulated_main_(int argc, char **argv)
409 simgrid::smpi::Process::init(&argc, &argv);
414 inline static int smpi_main_wrapper(int argc, char **argv){
415 int ret = smpi_simulated_main_(argc,argv);
417 XBT_WARN("SMPI process did not return 0. Return value : %d", ret);
418 smpi_process()->set_return_value(ret);
423 int __attribute__ ((weak)) main(int argc, char **argv)
425 return smpi_main(smpi_main_wrapper, argc, argv);
430 static void smpi_init_logs(){
432 /* Connect log categories. See xbt/log.c */
434 XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this function: xbt_log_appender_file.c depends on it
435 DO NOT connect this in XBT or so, or it will be useless to xbt_log_appender_file.c */
436 XBT_LOG_CONNECT(instr_smpi);
437 XBT_LOG_CONNECT(smpi_bench);
438 XBT_LOG_CONNECT(smpi_coll);
439 XBT_LOG_CONNECT(smpi_colls);
440 XBT_LOG_CONNECT(smpi_comm);
441 XBT_LOG_CONNECT(smpi_datatype);
442 XBT_LOG_CONNECT(smpi_dvfs);
443 XBT_LOG_CONNECT(smpi_group);
444 XBT_LOG_CONNECT(smpi_kernel);
445 XBT_LOG_CONNECT(smpi_mpi);
446 XBT_LOG_CONNECT(smpi_memory);
447 XBT_LOG_CONNECT(smpi_op);
448 XBT_LOG_CONNECT(smpi_pmpi);
449 XBT_LOG_CONNECT(smpi_request);
450 XBT_LOG_CONNECT(smpi_replay);
451 XBT_LOG_CONNECT(smpi_rma);
452 XBT_LOG_CONNECT(smpi_shared);
453 XBT_LOG_CONNECT(smpi_utils);
457 static void smpi_init_options(){
459 simgrid::smpi::Colls::set_collectives();
460 simgrid::smpi::Colls::smpi_coll_cleanup_callback=nullptr;
461 smpi_cpu_threshold = xbt_cfg_get_double("smpi/cpu-threshold");
462 smpi_host_speed = xbt_cfg_get_double("smpi/host-speed");
463 smpi_privatize_global_variables = xbt_cfg_get_boolean("smpi/privatize-global-variables");
464 if (smpi_cpu_threshold < 0)
465 smpi_cpu_threshold = DBL_MAX;
467 char* val = xbt_cfg_get_string("smpi/shared-malloc");
468 if (!strcasecmp(val, "yes") || !strcmp(val, "1") || !strcasecmp(val, "on") || !strcasecmp(val, "global")) {
469 smpi_cfg_shared_malloc = shmalloc_global;
470 } else if (!strcasecmp(val, "local")) {
471 smpi_cfg_shared_malloc = shmalloc_local;
472 } else if (!strcasecmp(val, "no") || !strcmp(val, "0") || !strcasecmp(val, "off")) {
473 smpi_cfg_shared_malloc = shmalloc_none;
475 xbt_die("Invalid value '%s' for option smpi/shared-malloc. Possible values: 'on' or 'global', 'local', 'off'",
480 int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
482 srand(SMPI_RAND_SEED);
484 if (getenv("SMPI_PRETEND_CC") != nullptr) {
485 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
486 * configuration tools */
491 TRACE_global_init(&argc, argv);
492 TRACE_add_start_function(TRACE_smpi_alloc);
493 TRACE_add_end_function(TRACE_smpi_release);
495 SIMIX_global_init(&argc, argv);
496 MSG_init(&argc,argv);
498 SMPI_switch_data_segment = &smpi_switch_data_segment;
502 // parse the platform file: get the host list
503 SIMIX_create_environment(argv[1]);
504 SIMIX_comm_set_copy_data_callback(smpi_comm_copy_data_callback);
505 SIMIX_function_register_default(realmain);
506 SIMIX_launch_application(argv[2]);
510 smpi_check_options();
512 if(smpi_privatize_global_variables)
513 smpi_initialize_global_memory_segments();
515 /* Clean IO before the run */
519 if (MC_is_active()) {
525 xbt_os_walltimer_stop(global_timer);
526 if (xbt_cfg_get_boolean("smpi/display-timing")){
527 double global_time = xbt_os_timer_elapsed(global_timer);
528 XBT_INFO("Simulated time: %g seconds. \n\n"
529 "The simulation took %g seconds (after parsing and platform setup)\n"
530 "%g seconds were actual computation of the application",
531 SIMIX_get_clock(), global_time , smpi_total_benched_time);
533 if (smpi_total_benched_time/global_time>=0.75)
534 XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
535 "You may want to use sampling functions or trace replay to reduce this.");
538 int count = smpi_process_count();
540 for (i = 0; i < count; i++) {
541 if(process_data[i]->return_value()!=0){
542 ret=process_data[i]->return_value();//return first non 0 value
546 smpi_global_destroy();
553 // This function can be called from extern file, to initialize logs, options, and processes of smpi
554 // without the need of smpirun
559 smpi_check_options();
560 if (TRACE_is_enabled() && TRACE_is_configured())
562 if(smpi_privatize_global_variables)
563 smpi_initialize_global_memory_segments();
566 void SMPI_finalize(){
567 smpi_global_destroy();
570 void smpi_mpi_init() {
571 if(smpi_init_sleep > 0)
572 simcall_process_sleep(smpi_init_sleep);
575 double smpi_mpi_wtime(){
577 if (smpi_process()->initialized() != 0 && smpi_process()->finalized() == 0 && smpi_process()->sampling() == 0) {
579 time = SIMIX_get_clock();
580 // to avoid deadlocks if used as a break condition, such as
581 // while (MPI_Wtime(...) < time_limit) {
584 // because the time will not normally advance when only calls to MPI_Wtime
585 // are made -> deadlock (MPI_Wtime never reaches the time limit)
586 if(smpi_wtime_sleep > 0)
587 simcall_process_sleep(smpi_wtime_sleep);
590 time = SIMIX_get_clock();