X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/6d822b05d062d8be5858745ee8bdba4208f42d00..c02acff354ef915c0c88943e73959ca5c081b524:/src/smpi/internals/smpi_global.cpp diff --git a/src/smpi/internals/smpi_global.cpp b/src/smpi/internals/smpi_global.cpp index 7910ac00f1..87b7b0781f 100644 --- a/src/smpi/internals/smpi_global.cpp +++ b/src/smpi/internals/smpi_global.cpp @@ -4,12 +4,13 @@ * under the terms of the license (GNU LGPL) which comes with this package. */ #include "mc/mc.h" -#include "simgrid/s4u/Engine.hpp" #include "simgrid/plugins/file_system.h" +#include "simgrid/s4u/Engine.hpp" #include "smpi_coll.hpp" +#include "smpi_config.hpp" #include "smpi_f2c.hpp" #include "smpi_host.hpp" -#include "smpi_config.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" @@ -20,7 +21,6 @@ #include #include /* split */ #include -#include #include #include #include /* intmax_t */ @@ -71,26 +71,28 @@ std::map std::unordered_map location2speedup; static int smpi_exit_status = 0; -extern double smpi_total_benched_time; xbt_os_timer_t global_timer; static std::vector privatize_libs_paths; -/** - * Setting MPI_COMM_WORLD to MPI_COMM_UNINITIALIZED (it's a variable) - * is important because the implementation of MPI_Comm checks - * "this == MPI_COMM_UNINITIALIZED"? If yes, it uses smpi_process()->comm_world() - * instead of "this". - * This is basically how we only have one global variable but all processes have - * different communicators (the one their SMPI instance uses). - * - * See smpi_comm.cpp and the functions therein for details. - */ -MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED; + // No instance gets manually created; check also the smpirun.in script as // this default name is used there as well (when the tag is generated). static const std::string smpi_default_instance_name("smpirun"); static simgrid::config::Flag smpi_init_sleep( "smpi/init", "Time to inject inside a call to MPI_Init", 0.0); +static simgrid::config::Flag + smpi_hostfile("smpi/hostfile", + "Classical MPI hostfile containing list of machines to dispatch " + "the processes, one per line", + ""); + +static simgrid::config::Flag smpi_replay("smpi/replay", + "Replay a trace instead of executing the application", ""); + +static simgrid::config::Flag smpi_np("smpi/np", "Number of processes to be created", 0); + +static simgrid::config::Flag smpi_map("smpi/map", "Display the mapping between nodes and processes", 0); + void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*, size_t) = &smpi_comm_copy_buffer_callback; @@ -195,21 +197,15 @@ void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, v auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks); check_blocks(private_blocks, buff_size); void* tmpbuff=buff; - if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) && - (static_cast(buff) >= smpi_data_exe_start) && - (static_cast(buff) < smpi_data_exe_start + smpi_data_exe_size)) { + if (smpi_switch_data_segment(comm->src_actor_->get_iface(), buff)) { XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !"); - smpi_switch_data_segment(comm->src_actor_->get_iface()); tmpbuff = xbt_malloc(buff_size); memcpy_private(tmpbuff, buff, private_blocks); } - if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) && - ((char*)comm->dst_buff_ >= smpi_data_exe_start) && - ((char*)comm->dst_buff_ < smpi_data_exe_start + smpi_data_exe_size)) { + if (smpi_switch_data_segment(comm->dst_actor_->get_iface(), comm->dst_buff_)) XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment"); - smpi_switch_data_segment(comm->dst_actor_->get_iface()); - } + XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_); memcpy_private(comm->dst_buff_, tmpbuff, private_blocks); @@ -362,8 +358,8 @@ static void smpi_copy_file(const std::string& src, const std::string& target, of { int fdin = open(src.c_str(), O_RDONLY); xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str()); - XBT_ATTRIB_UNUSED int unlink_status = unlink(target.c_str()); - xbt_assert(unlink_status == 0 || errno == ENOENT, "Failed to unlink file %s: %s", target.c_str(), strerror(errno)); + xbt_assert(unlink(target.c_str()) == 0 || errno == ENOENT, "Failed to unlink file %s: %s", target.c_str(), + strerror(errno)); int fdout = open(target.c_str(), O_CREAT | O_RDWR | O_EXCL, S_IRWXU); xbt_assert(fdout >= 0, "Cannot write into %s: %s", target.c_str(), strerror(errno)); @@ -374,10 +370,10 @@ static void smpi_copy_file(const std::string& src, const std::string& target, of close(fdin); close(fdout); return; - } else if (sent_size != -1 || errno != ENOSYS) { - xbt_die("Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(), - sent_size, static_cast(fdin_size), errno, strerror(errno)); } + xbt_assert(sent_size == -1 && errno == ENOSYS, + "Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(), + sent_size, static_cast(fdin_size), errno, strerror(errno)); #endif // If this point is reached, sendfile() actually is not available. Copy file by hand. std::vector buf(1024 * 1024 * 4); @@ -430,14 +426,13 @@ static void smpi_init_privatization_dlopen(const std::string& executable) for (auto const& libname : privatize_libs) { // load the library once to add it to the local libs, to get the absolute path void* libhandle = dlopen(libname.c_str(), RTLD_LAZY); - xbt_assert(libhandle != nullptr, - "Cannot dlopen %s - check your settings in smpi/privatize-libs", libname.c_str()); + xbt_assert(libhandle != nullptr, "Cannot dlopen %s - check your settings in smpi/privatize-libs", + libname.c_str()); // get library name from path std::string fullpath = libname; #if not defined(__APPLE__) && not defined(__HAIKU__) - XBT_ATTRIB_UNUSED int dl_iterate_res = dl_iterate_phdr(visit_libs, &fullpath); - xbt_assert(dl_iterate_res != 0, "Can't find a linked %s - check your settings in smpi/privatize-libs", - fullpath.c_str()); + xbt_assert(dl_iterate_phdr(visit_libs, &fullpath) != 0, + "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath.c_str()); XBT_DEBUG("Extra lib to privatize '%s' found", fullpath.c_str()); #else xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku"); @@ -501,7 +496,7 @@ static void smpi_init_privatization_dlopen(const std::string& executable) dlerror(), saved_errno, strerror(saved_errno)); smpi_entry_point_type entry_point = smpi_resolve_function(handle); - xbt_assert(entry_point, "Could not resolve entry point"); + xbt_assert(entry_point, "Could not resolve entry point. Does your program contain a main() function?"); smpi_run_entry_point(entry_point, executable, args); }); }); @@ -524,8 +519,16 @@ static void smpi_init_privatization_no_dlopen(const std::string& executable) // Execute the same entry point for each simulated process: simgrid::s4u::Engine::get_instance()->register_default([entry_point, executable](std::vector args) { - return std::function( - [entry_point, executable, args] { smpi_run_entry_point(entry_point, executable, args); }); + return std::function([entry_point, executable, args] { + if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) { + simgrid::smpi::ActorExt* ext = smpi_process(); + /* Now using the segment index of this process */ + ext->set_privatized_region(smpi_init_global_memory_segment_process()); + /* Done at the process's creation */ + smpi_switch_data_segment(simgrid::s4u::Actor::self()); + } + smpi_run_entry_point(entry_point, executable, args); + }); }); } @@ -536,9 +539,8 @@ int smpi_main(const char* executable, int argc, char* argv[]) * configuration tools */ return 0; } - - SMPI_switch_data_segment = &smpi_switch_data_segment; - smpi_init_options(); + + smpi_init_options(true); simgrid::instr::init(); SIMIX_global_init(&argc, argv); @@ -547,7 +549,7 @@ int smpi_main(const char* executable, int argc, char* argv[]) sg_storage_file_system_init(); // parse the platform file: get the host list engine->load_platform(argv[1]); - SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback); + simgrid::kernel::activity::CommImpl::set_copy_data_callback(smpi_comm_copy_buffer_callback); if (smpi_cfg_privatization() == SmpiPrivStrategies::DLOPEN) smpi_init_privatization_dlopen(executable); @@ -559,13 +561,8 @@ int smpi_main(const char* executable, int argc, char* argv[]) SMPI_init(); - /* This is a ... heavy way to count the MPI ranks */ - int rank_counts = 0; - simgrid::s4u::Actor::on_creation.connect([&rank_counts](const simgrid::s4u::Actor& actor) { - if (not actor.is_daemon()) - rank_counts++; - }); - engine->load_deployment(argv[2]); + int rank_counts = smpi_deployment_smpirun(engine, smpi_hostfile.get(), smpi_np.get(), smpi_replay.get(), + smpi_map.get(), argc - 2, argv + 2); SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, rank_counts); MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name); @@ -577,20 +574,10 @@ int smpi_main(const char* executable, int argc, char* argv[]) if (MC_is_active()) { MC_run(); } else { - SIMIX_run(); + simgrid::kernel::EngineImpl::get_instance()->run(); xbt_os_walltimer_stop(global_timer); - if (simgrid::config::get_value("smpi/display-timing")) { - double global_time = xbt_os_timer_elapsed(global_timer); - XBT_INFO("Simulated time: %g seconds. \n\n" - "The simulation took %g seconds (after parsing and platform setup)\n" - "%g seconds were actual computation of the application", - SIMIX_get_clock(), global_time , smpi_total_benched_time); - - if (smpi_total_benched_time/global_time>=0.75) - XBT_INFO("More than 75%% of the time was spent inside the application code.\n" - "You may want to use sampling functions or trace replay to reduce this."); - } + simgrid::smpi::utils::print_time_analysis(xbt_os_timer_elapsed(global_timer)); } SMPI_finalize(); @@ -634,23 +621,8 @@ void SMPI_finalize() if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) smpi_destroy_global_memory_segments(); - if (simgrid::smpi::F2C::lookup() != nullptr && - simgrid::smpi::F2C::lookup()->size() > simgrid::smpi::F2C::get_num_default_handles()) { - XBT_WARN("Probable Leaks in your code: SMPI detected %zu unfreed MPI handles : " - "display types and addresses (n max) with --cfg=smpi/list-leaks:n.\n" - "Running smpirun with -wrapper \"valgrind --leak-check=full\" can provide more information", - simgrid::smpi::F2C::lookup()->size() - simgrid::smpi::F2C::get_num_default_handles()); - int n = simgrid::config::get_value("smpi/list-leaks"); - for (auto const& p : *simgrid::smpi::F2C::lookup()) { - static int printed = 0; - if (printed >= n) - break; - if (p.first >= simgrid::smpi::F2C::get_num_default_handles()) { - XBT_WARN("Leak %p of type %s", p.second, boost::core::demangle(typeid(*(p.second)).name()).c_str()); - printed++; - } - } - } + + simgrid::smpi::utils::print_memory_analysis(); } void smpi_mpi_init() { @@ -661,4 +633,5 @@ void smpi_mpi_init() { void SMPI_thread_create() { TRACE_smpi_init(simgrid::s4u::this_actor::get_pid(), __func__); + smpi_process()->mark_as_initialized(); }