* under the terms of the license (GNU LGPL) which comes with this package. */
#include "mc/mc.h"
+#include "simgrid/Exception.hpp"
#include "simgrid/plugins/file_system.h"
#include "simgrid/s4u/Engine.hpp"
#include "smpi_coll.hpp"
#include "smpi_host.hpp"
#include "src/kernel/EngineImpl.hpp"
#include "src/kernel/activity/CommImpl.hpp"
-#include "src/simix/smx_private.hpp"
#include "src/smpi/include/smpi_actor.hpp"
#include "xbt/config.hpp"
#include "xbt/file.hpp"
// No instance gets manually created; check also the smpirun.in script as
// this default name is used there as well (when the <actor> tag is generated).
static const std::string smpi_default_instance_name("smpirun");
-static simgrid::config::Flag<double> smpi_init_sleep(
- "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
static simgrid::config::Flag<std::string>
smpi_hostfile("smpi/hostfile",
// the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
// and the (computed) event_set.
- if (not smpi_cfg_papi_events_file().empty()) {
- if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
- XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?"
- " Expected version is %u", PAPI_VER_CURRENT);
-
- using Tokenizer = boost::tokenizer<boost::char_separator<char>>;
- boost::char_separator<char> separator_units(";");
- std::string str = smpi_cfg_papi_events_file();
- Tokenizer tokens(str, separator_units);
-
- // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it.
- // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units".
- for (auto const& unit_it : tokens) {
- boost::char_separator<char> separator_events(":");
- Tokenizer event_tokens(unit_it, separator_events);
-
- int event_set = PAPI_NULL;
- if (PAPI_create_eventset(&event_set) != PAPI_OK) {
- // TODO: Should this let the whole simulation die?
- XBT_CRITICAL("Could not create PAPI event set during init.");
- }
+ if (smpi_cfg_papi_events_file().empty())
+ return;
- // NOTE: We cannot use a map here, as we must obey the order of the counters
- // This is important for PAPI: We need to map the values of counters back to the event_names (so, when PAPI_read()
- // has finished)!
- papi_counter_t counters2values;
-
- // Iterate over all counters that were specified for this specific unit.
- // Note that we need to remove the name of the unit (that could also be the "default" value), which always comes
- // first. Hence, we start at ++(events.begin())!
- for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) {
- int event_code = PAPI_NULL;
- auto* event_name = const_cast<char*>((*events_it).c_str());
- if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) {
- XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
- continue;
- }
- if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
- XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
- continue;
- }
- XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
+ if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT) {
+ XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked? Expected version is %u",
+ PAPI_VER_CURRENT);
+ return;
+ }
- counters2values.emplace_back(*events_it, 0LL);
- }
+ using Tokenizer = boost::tokenizer<boost::char_separator<char>>;
+ boost::char_separator<char> separator_units(";");
+ std::string str = smpi_cfg_papi_events_file();
+ Tokenizer tokens(str, separator_units);
+
+ // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it.
+ // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units".
+ for (auto const& unit_it : tokens) {
+ boost::char_separator<char> separator_events(":");
+ Tokenizer event_tokens(unit_it, separator_events);
+
+ int event_set = PAPI_NULL;
+ if (PAPI_create_eventset(&event_set) != PAPI_OK) {
+ // TODO: Should this let the whole simulation die?
+ XBT_CRITICAL("Could not create PAPI event set during init.");
+ break;
+ }
- std::string unit_name = *(event_tokens.begin());
- papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
+ // NOTE: We cannot use a map here, as we must obey the order of the counters
+ // This is important for PAPI: We need to map the values of counters back to the event_names (so, when PAPI_read()
+ // has finished)!
+ papi_counter_t counters2values;
+
+ // Iterate over all counters that were specified for this specific unit.
+ // Note that we need to remove the name of the unit (that could also be the "default" value), which always comes
+ // first. Hence, we start at ++(events.begin())!
+ for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) {
+ int event_code = PAPI_NULL;
+ auto* event_name = const_cast<char*>((*events_it).c_str());
+ if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) {
+ XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
+ continue;
+ }
+ if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
+ XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
+ continue;
+ }
+ XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
- units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
+ counters2values.emplace_back(*events_it, 0LL);
}
+
+ std::string unit_name = *(event_tokens.begin());
+ papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
+
+ units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
}
#endif
}
using smpi_fortran_entry_point_type = void (*)();
template <typename F>
-static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path, std::vector<std::string> args)
+static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path,
+ const std::vector<std::string>& args)
{
// copy C strings, we need them writable
- auto* args4argv = new std::vector<char*>(args.size());
- std::transform(begin(args), end(args), begin(*args4argv), [](const std::string& s) { return xbt_strdup(s.c_str()); });
+ std::vector<char*> args4argv(args.size());
+ std::transform(begin(args) + 1, end(args), begin(args4argv) + 1,
+ [](const std::string& s) { return xbt_strdup(s.c_str()); });
// set argv[0] to executable_path
- xbt_free((*args4argv)[0]);
- (*args4argv)[0] = xbt_strdup(executable_path.c_str());
+ args4argv[0] = xbt_strdup(executable_path.c_str());
+ // add the final NULL
+ args4argv.push_back(nullptr);
-#if !SMPI_IFORT
// take a copy of args4argv to keep reference of the allocated strings
- const std::vector<char*> args2str(*args4argv);
-#endif
- int argc = args4argv->size();
- args4argv->push_back(nullptr);
- char** argv = args4argv->data();
-
-#if SMPI_IFORT
- for_rtl_init_ (&argc, argv);
-#elif SMPI_FLANG
- __io_set_argc(argc);
- __io_set_argv(argv);
-#elif SMPI_GFORTRAN
- _gfortran_set_args(argc, argv);
-#endif
- int res = entry_point(argc, argv);
+ const std::vector<char*> args2str(args4argv);
+
+ try {
+ int argc = static_cast<int>(args4argv.size() - 1);
+ char** argv = args4argv.data();
+ int res = entry_point(argc, argv);
+ if (res != 0) {
+ XBT_WARN("SMPI process did not return 0. Return value : %d", res);
+ if (smpi_exit_status == 0)
+ smpi_exit_status = res;
+ }
+ } catch (simgrid::ForcefulKillException const& e) {
+ XBT_DEBUG("Caught a ForcefulKillException: %s", e.what());
+ }
-#if SMPI_IFORT
- for_rtl_finish_ ();
-#else
for (char* s : args2str)
xbt_free(s);
- delete args4argv;
-#endif
- if (res != 0){
- XBT_WARN("SMPI process did not return 0. Return value : %d", res);
- if (smpi_exit_status == 0)
- smpi_exit_status = res;
- }
return 0;
}
-
-// TODO, remove the number of functions involved here
static smpi_entry_point_type smpi_resolve_function(void* handle)
{
auto* entry_point_fortran = reinterpret_cast<smpi_fortran_entry_point_type>(dlsym(handle, "user_main_"));
while (ssize_t got = read(fdin, buf.data(), buf.size())) {
if (got == -1) {
xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str());
- } else {
- const unsigned char* p = buf.data();
- ssize_t todo = got;
- while (ssize_t done = write(fdout, p, todo)) {
- if (done == -1) {
- xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str());
- } else {
- p += done;
- todo -= done;
- }
+ continue;
+ }
+ const unsigned char* p = buf.data();
+ ssize_t todo = got;
+ while (ssize_t done = write(fdout, p, todo)) {
+ if (done == -1) {
+ xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str());
+ continue;
}
+ p += done;
+ todo -= done;
}
}
close(fdin);
}
simgrid::s4u::Engine::get_instance()->register_default([executable, fdin_size](std::vector<std::string> args) {
- return std::function<void()>([executable, fdin_size, args] {
+ return simgrid::kernel::actor::ActorCode([executable, fdin_size, args = std::move(args)] {
static std::size_t rank = 0;
// Copy the dynamic library:
simgrid::xbt::Path path(executable);
// Copy the dynamic library, the new name must be the same length as the old one
// just replace the name with 7 digits for the rank and the rest of the name.
- auto pad = std::min<unsigned>(7, libname.length());
+ auto pad = std::min<size_t>(7, libname.length());
std::string target_libname = std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
std::string target_lib = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" + target_libname;
target_libs.push_back(target_lib);
// Execute the same entry point for each simulated process:
simgrid::s4u::Engine::get_instance()->register_default([entry_point, executable](std::vector<std::string> args) {
- return std::function<void()>([entry_point, executable, args] {
+ return simgrid::kernel::actor::ActorCode([entry_point, executable, args = std::move(args)] {
if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
simgrid::smpi::ActorExt* ext = smpi_process();
/* Now using the segment index of this process */
}
smpi_init_options_internal(true);
- simgrid::instr::init();
- SIMIX_global_init(&argc, argv);
-
- auto engine = simgrid::s4u::Engine::get_instance();
+ simgrid::s4u::Engine engine(&argc, argv);
sg_storage_file_system_init();
// parse the platform file: get the host list
- engine->load_platform(argv[1]);
- simgrid::kernel::activity::CommImpl::set_copy_data_callback(smpi_comm_copy_buffer_callback);
+ engine.load_platform(argv[1]);
+ engine.set_default_comm_data_copy_callback(smpi_comm_copy_buffer_callback);
if (smpi_cfg_privatization() == SmpiPrivStrategies::DLOPEN)
smpi_init_privatization_dlopen(executable);
simgrid::smpi::colls::set_collectives();
simgrid::smpi::colls::smpi_coll_cleanup_callback = nullptr;
-
+
+ std::vector<char*> args4argv(argv + 1, argv + argc + 1); // last element is NULL
+ args4argv[0] = xbt_strdup(executable);
+ int real_argc = argc - 1;
+ char** real_argv = args4argv.data();
+
+ // Setup argc/argv for the Fortran run-time environment
+#if SMPI_IFORT
+ for_rtl_init_(&real_argc, real_argv);
+#elif SMPI_FLANG
+ __io_set_argc(real_argc);
+ __io_set_argv(real_argv);
+#elif SMPI_GFORTRAN
+ _gfortran_set_args(real_argc, real_argv);
+#endif
+
SMPI_init();
- int rank_counts = smpi_deployment_smpirun(engine, smpi_hostfile.get(), smpi_np.get(), smpi_replay.get(),
- smpi_map.get(), argc - 2, argv + 2);
+ const std::vector<const char*> args(real_argv + 1, real_argv + real_argc);
+ int rank_counts =
+ smpi_deployment_smpirun(&engine, smpi_hostfile.get(), smpi_np.get(), smpi_replay.get(), smpi_map.get(), args);
SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, rank_counts);
MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name);
if (MC_is_active()) {
MC_run();
} else {
- simgrid::kernel::EngineImpl::get_instance()->run();
+ engine.get_impl()->run();
xbt_os_walltimer_stop(global_timer);
simgrid::smpi::utils::print_time_analysis(xbt_os_timer_elapsed(global_timer));
}
SMPI_finalize();
+#if SMPI_IFORT
+ for_rtl_finish_();
+#endif
+ xbt_free(args4argv[0]);
+
return smpi_exit_status;
}
void smpi_mpi_init() {
smpi_init_fortran_types();
- if(smpi_init_sleep > 0)
- simgrid::s4u::this_actor::sleep_for(smpi_init_sleep);
+ if(_smpi_init_sleep > 0)
+ simgrid::s4u::this_actor::sleep_for(_smpi_init_sleep);
+ if (not MC_is_active()) {
+ smpi_deployment_startup_barrier(smpi_process()->get_instance_id());
+ }
}
void SMPI_thread_create() {
TRACE_smpi_init(simgrid::s4u::this_actor::get_pid(), __func__);
smpi_process()->mark_as_initialized();
}
+
+void smpi_exit(int res){
+ if(res != 0){
+ XBT_WARN("SMPI process did not return 0. Return value : %d", res);
+ smpi_exit_status = res;
+ }
+ simgrid::s4u::this_actor::exit();
+ THROW_IMPOSSIBLE;
+}
