-/* Copyright (c) 2007-2018. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2007-2019. The SimGrid Team. All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
* under the terms of the license (GNU LGPL) which comes with this package. */
-#include "SmpiHost.hpp"
#include "mc/mc.h"
-#include "private.hpp"
#include "simgrid/s4u/Engine.hpp"
-#include "simgrid/s4u/Host.hpp"
-#include "simgrid/s4u/Mailbox.hpp"
-#include "simgrid/s4u/forward.hpp"
+#include "simgrid/plugins/file_system.h"
#include "smpi_coll.hpp"
-#include "smpi_comm.hpp"
-#include "smpi_group.hpp"
-#include "smpi_info.hpp"
-#include "smpi_process.hpp"
-#include "src/msg/msg_private.hpp"
+#include "smpi_f2c.hpp"
+#include "smpi_host.hpp"
+#include "src/kernel/activity/CommImpl.hpp"
#include "src/simix/smx_private.hpp"
-#include "src/surf/surf_interface.hpp"
+#include "src/smpi/include/smpi_actor.hpp"
#include "xbt/config.hpp"
+#include <algorithm>
+#include <boost/algorithm/string.hpp> /* trim_right / trim_left */
+#include <boost/tokenizer.hpp>
#include <cfloat> /* DBL_MAX */
+#include <cinttypes>
+#include <cstdint> /* intmax_t */
#include <dlfcn.h>
#include <fcntl.h>
#include <fstream>
#include <sys/stat.h>
-#if HAVE_SENDFILE
+#if SIMGRID_HAVE_MC
+#include "src/mc/mc_config.hpp"
+#endif
+
+#if SG_HAVE_SENDFILE
#include <sys/sendfile.h>
#endif
+#if HAVE_PAPI
+#include "papi.h"
+#endif
+
+#if not defined(__APPLE__) && not defined(__HAIKU__)
+#include <link.h>
+#endif
+
+#if defined(__APPLE__)
+# include <AvailabilityMacros.h>
+# ifndef MAC_OS_X_VERSION_10_12
+# define MAC_OS_X_VERSION_10_12 101200
+# endif
+constexpr bool HAVE_WORKING_MMAP = (MAC_OS_X_VERSION_MIN_REQUIRED >= MAC_OS_X_VERSION_10_12);
+#elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__sun) || defined(__HAIKU__)
+constexpr bool HAVE_WORKING_MMAP = false;
+#else
+constexpr bool HAVE_WORKING_MMAP = true;
+#endif
+
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)");
-#include <boost/tokenizer.hpp>
-#include <boost/algorithm/string.hpp> /* trim_right / trim_left */
-#ifndef RTLD_DEEPBIND
+#if SMPI_IFORT
+ extern "C" void for_rtl_init_ (int *, char **);
+ extern "C" void for_rtl_finish_ ();
+#elif SMPI_FLANG
+ extern "C" void __io_set_argc(int);
+ extern "C" void __io_set_argv(char **);
+#elif SMPI_GFORTRAN
+ extern "C" void _gfortran_set_args(int, char **);
+#endif
+
/* RTLD_DEEPBIND is a bad idea of GNU ld that obviously does not exist on other platforms
* See https://www.akkadia.org/drepper/dsohowto.pdf
* and https://lists.freebsd.org/pipermail/freebsd-current/2016-March/060284.html
*/
-#define RTLD_DEEPBIND 0
+#if !RTLD_DEEPBIND || HAVE_SANITIZER_ADDRESS || HAVE_SANITIZER_THREAD
+#define WANT_RTLD_DEEPBIND 0
+#else
+#define WANT_RTLD_DEEPBIND RTLD_DEEPBIND
#endif
#if HAVE_PAPI
-#include "papi.h"
-const char* papi_default_config_name = "default";
-
-struct papi_process_data {
- papi_counter_t counter_data;
- int event_set;
-};
-
+std::string papi_default_config_name = "default";
+std::map</* computation unit name */ std::string, papi_process_data> units2papi_setup;
#endif
-using simgrid::s4u::Actor;
-using simgrid::s4u::ActorPtr;
+
std::unordered_map<std::string, double> location2speedup;
-static std::map</*process_id*/ ActorPtr, simgrid::smpi::Process*> process_data;
-int process_count = 0;
+static std::map</*process_id*/ simgrid::s4u::Actor const*, simgrid::smpi::ActorExt*> process_data;
static int smpi_exit_status = 0;
-int smpi_universe_size = 0;
extern double smpi_total_benched_time;
xbt_os_timer_t global_timer;
+static std::vector<std::string> privatize_libs_paths;
/**
* Setting MPI_COMM_WORLD to MPI_COMM_UNINITIALIZED (it's a variable)
* is important because the implementation of MPI_Comm checks
* See smpi_comm.cpp and the functions therein for details.
*/
MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
-MPI_Errhandler *MPI_ERRORS_RETURN = nullptr;
-MPI_Errhandler *MPI_ERRORS_ARE_FATAL = nullptr;
-MPI_Errhandler *MPI_ERRHANDLER_NULL = nullptr;
// 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_wtime_sleep(
- "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0);
static simgrid::config::Flag<double> smpi_init_sleep(
"smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
-void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t) = &smpi_comm_copy_buffer_callback;
+void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*,
+ size_t) = &smpi_comm_copy_buffer_callback;
-int smpi_process_count()
+simgrid::smpi::ActorExt* smpi_process()
{
- return process_count;
-}
+ simgrid::s4u::ActorPtr me = simgrid::s4u::Actor::self();
-simgrid::smpi::Process* smpi_process()
-{
- ActorPtr me = Actor::self();
if (me == nullptr) // This happens sometimes (eg, when linking against NS3 because it pulls openMPI...)
return nullptr;
- simgrid::msg::ActorExt* msgExt = static_cast<simgrid::msg::ActorExt*>(me->getImpl()->userdata);
- return static_cast<simgrid::smpi::Process*>(msgExt->data);
+
+ return process_data.at(me.get());
}
-simgrid::smpi::Process* smpi_process_remote(ActorPtr actor)
+simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor)
{
- return process_data.at(actor);
+ if (actor.get() == nullptr)
+ return nullptr;
+ return process_data.at(actor.get());
}
MPI_Comm smpi_process_comm_self(){
return smpi_process()->comm_self();
}
-void smpi_process_init(int *argc, char ***argv){
- simgrid::smpi::Process::init(argc, argv);
-}
-
-int smpi_process_index(){
- return simgrid::s4u::Actor::self()->getPid();
+MPI_Info smpi_process_info_env(){
+ return smpi_process()->info_env();
}
void * smpi_process_get_user_data(){
- return smpi_process()->get_user_data();
+ return simgrid::s4u::Actor::self()->get_impl()->get_user_data();
}
void smpi_process_set_user_data(void *data){
- return smpi_process()->set_user_data(data);
-}
-
-
-int smpi_global_size()
-{
- char *value = getenv("SMPI_GLOBAL_SIZE");
- xbt_assert(value,"Please set env var SMPI_GLOBAL_SIZE to the expected number of processes.");
-
- return xbt_str_parse_int(value, "SMPI_GLOBAL_SIZE contains a non-numerical value: %s");
+ simgrid::s4u::Actor::self()->get_impl()->set_user_data(data);
}
void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
{
- smpi_comm_copy_data_callback = callback;
+ static void (*saved_callback)(smx_activity_t, void*, size_t);
+ saved_callback = callback;
+ smpi_comm_copy_data_callback = [](simgrid::kernel::activity::CommImpl* comm, void* buff, size_t size) {
+ saved_callback(smx_activity_t(comm), buff, size);
+ };
}
-static void print(std::vector<std::pair<size_t, size_t>> vec) {
- std::fprintf(stderr, "{");
- for (auto const& elt : vec) {
- std::fprintf(stderr, "(0x%zx, 0x%zx),", elt.first, elt.second);
- }
- std::fprintf(stderr, "}\n");
-}
static void memcpy_private(void* dest, const void* src, std::vector<std::pair<size_t, size_t>>& private_blocks)
{
for (auto const& block : private_blocks)
xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
}
-void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size)
+void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff, size_t buff_size)
{
- simgrid::kernel::activity::CommImplPtr comm =
- boost::dynamic_pointer_cast<simgrid::kernel::activity::CommImpl>(synchro);
- int src_shared = 0;
- int dst_shared = 0;
size_t src_offset = 0;
size_t dst_offset = 0;
std::vector<std::pair<size_t, size_t>> src_private_blocks;
std::vector<std::pair<size_t, size_t>> dst_private_blocks;
XBT_DEBUG("Copy the data over");
- if((src_shared=smpi_is_shared(buff, src_private_blocks, &src_offset))) {
+ if(smpi_is_shared(buff, src_private_blocks, &src_offset)) {
XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff);
src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
}
src_private_blocks.clear();
src_private_blocks.push_back(std::make_pair(0, buff_size));
}
- if((dst_shared=smpi_is_shared((char*)comm->dst_buff, dst_private_blocks, &dst_offset))) {
- XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff);
+ if (smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset)) {
+ XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff_);
dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
}
else {
auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
check_blocks(private_blocks, buff_size);
void* tmpbuff=buff;
- if ((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && (static_cast<char*>(buff) >= smpi_data_exe_start) &&
+ if ((smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) &&
+ (static_cast<char*>(buff) >= smpi_data_exe_start) &&
(static_cast<char*>(buff) < smpi_data_exe_start + smpi_data_exe_size)) {
XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
- smpi_switch_data_segment(comm->src_proc->iface());
+ smpi_switch_data_segment(comm->src_actor_->iface());
tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
memcpy_private(tmpbuff, buff, private_blocks);
}
- if ((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_data_exe_start) &&
- ((char*)comm->dst_buff < smpi_data_exe_start + smpi_data_exe_size)) {
+ if ((smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) &&
+ ((char*)comm->dst_buff_ >= smpi_data_exe_start) &&
+ ((char*)comm->dst_buff_ < smpi_data_exe_start + smpi_data_exe_size)) {
XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
- smpi_switch_data_segment(comm->dst_proc->iface());
+ smpi_switch_data_segment(comm->dst_actor_->iface());
}
- XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
- memcpy_private(comm->dst_buff, tmpbuff, private_blocks);
+ XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_);
+ memcpy_private(comm->dst_buff_, tmpbuff, private_blocks);
- if (comm->detached) {
+ if (comm->detached()) {
// if this is a detached send, the source buffer was duplicated by SMPI
// sender to make the original buffer available to the application ASAP
xbt_free(buff);
//It seems that the request is used after the call there this should be free somewhere else but where???
//xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
- comm->src_buff = nullptr;
+ comm->src_buff_ = nullptr;
}
if (tmpbuff != buff)
xbt_free(tmpbuff);
}
-void smpi_comm_null_copy_buffer_callback(smx_activity_t comm, void *buff, size_t buff_size)
+void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, void*, size_t)
{
/* nothing done in this version */
}
-static void smpi_check_options(){
- //check correctness of MPI parameters
-
- xbt_assert(xbt_cfg_get_int("smpi/async-small-thresh") <= xbt_cfg_get_int("smpi/send-is-detached-thresh"));
+static void smpi_check_options()
+{
+#if SIMGRID_HAVE_MC
+ if (MC_is_active()) {
+ if (_sg_mc_buffering == "zero")
+ simgrid::config::set_value<int>("smpi/send-is-detached-thresh", 0);
+ else if (_sg_mc_buffering == "infty")
+ simgrid::config::set_value<int>("smpi/send-is-detached-thresh", INT_MAX);
+ else
+ THROW_IMPOSSIBLE;
+ }
+#endif
- if (xbt_cfg_is_default_value("smpi/host-speed")) {
- XBT_INFO("You did not set the power of the host running the simulation. "
- "The timings will certainly not be accurate. "
- "Use the option \"--cfg=smpi/host-speed:<flops>\" to set its value."
- "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
- }
+ xbt_assert(simgrid::config::get_value<int>("smpi/async-small-thresh") <=
+ simgrid::config::get_value<int>("smpi/send-is-detached-thresh"),
+ "smpi/async-small-thresh (=%d) should be smaller or equal to smpi/send-is-detached-thresh (=%d)",
+ simgrid::config::get_value<int>("smpi/async-small-thresh"),
+ simgrid::config::get_value<int>("smpi/send-is-detached-thresh"));
+
+ if (simgrid::config::is_default("smpi/host-speed") && not MC_is_active()) {
+ XBT_INFO("You did not set the power of the host running the simulation. "
+ "The timings will certainly not be accurate. "
+ "Use the option \"--cfg=smpi/host-speed:<flops>\" to set its value. "
+ "Check "
+ "https://simgrid.org/doc/latest/Configuring_SimGrid.html#automatic-benchmarking-of-smpi-code for more "
+ "information.");
+ }
- xbt_assert(xbt_cfg_get_double("smpi/cpu-threshold") >=0,
- "The 'smpi/cpu-threshold' option cannot have negative values [anymore]. If you want to discard "
- "the simulation of any computation, please use 'smpi/simulate-computation:no' instead.");
+ xbt_assert(simgrid::config::get_value<double>("smpi/cpu-threshold") >= 0,
+ "The 'smpi/cpu-threshold' option cannot have negative values [anymore]. If you want to discard "
+ "the simulation of any computation, please use 'smpi/simulate-computation:no' instead.");
}
int smpi_enabled() {
- return not process_data.empty();
+ return MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED;
}
-void smpi_global_init()
+static void smpi_init_papi()
{
- if (not MC_is_active()) {
- global_timer = xbt_os_timer_new();
- xbt_os_walltimer_start(global_timer);
- }
-
- std::string filename = xbt_cfg_get_string("smpi/comp-adjustment-file");
- if (not filename.empty()) {
- std::ifstream fstream(filename);
- if (not fstream.is_open()) {
- xbt_die("Could not open file %s. Does it exist?", filename.c_str());
- }
-
- std::string line;
- typedef boost::tokenizer< boost::escaped_list_separator<char>> Tokenizer;
- std::getline(fstream, line); // Skip the header line
- while (std::getline(fstream, line)) {
- Tokenizer tok(line);
- Tokenizer::iterator it = tok.begin();
- Tokenizer::iterator end = std::next(tok.begin());
-
- std::string location = *it;
- boost::trim(location);
- location2speedup.insert(std::pair<std::string, double>(location, std::stod(*end)));
- }
- }
-
#if HAVE_PAPI
// This map holds for each computation unit (such as "default" or "process1" etc.)
// the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
// and the (computed) event_set.
- std::map</* computation unit name */ std::string, papi_process_data> units2papi_setup;
- if (not xbt_cfg_get_string("smpi/papi-events").empty()) {
+ if (not simgrid::config::get_value<std::string>("smpi/papi-events").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 %i",
- PAPI_VER_CURRENT);
+ " Expected version is %u", PAPI_VER_CURRENT);
typedef boost::tokenizer<boost::char_separator<char>> Tokenizer;
boost::char_separator<char> separator_units(";");
- std::string str = xbt_cfg_get_string("smpi/papi-events");
+ std::string str = simgrid::config::get_value<std::string>("smpi/papi-events");
Tokenizer tokens(str, separator_units);
// Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it.
int event_code = PAPI_NULL;
char* event_name = const_cast<char*>((*events_it).c_str());
- if (PAPI_event_name_to_code(event_name, &event_code) == PAPI_OK) {
- if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
- XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
- continue;
- } else {
- XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
- }
- } else {
+ 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);
counters2values.push_back(
// We cannot just pass *events_it, as this is of type const basic_string
#endif
}
-void smpi_global_destroy()
-{
- smpi_bench_destroy();
- smpi_shared_destroy();
- smpi_deployment_cleanup_instances();
-
- if (simgrid::smpi::Colls::smpi_coll_cleanup_callback != nullptr)
- simgrid::smpi::Colls::smpi_coll_cleanup_callback();
-
- MPI_COMM_WORLD = MPI_COMM_NULL;
-
- if (not MC_is_active()) {
- xbt_os_timer_free(global_timer);
- }
-
- if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
- smpi_destroy_global_memory_segments();
- smpi_free_static();
-}
-
static void smpi_init_options(){
// return if already called
if (smpi_cpu_threshold > -1)
return;
simgrid::smpi::Colls::set_collectives();
simgrid::smpi::Colls::smpi_coll_cleanup_callback = nullptr;
- smpi_cpu_threshold = xbt_cfg_get_double("smpi/cpu-threshold");
- smpi_host_speed = xbt_cfg_get_double("smpi/host-speed");
- std::string smpi_privatize_option = xbt_cfg_get_string("smpi/privatization");
+ smpi_cpu_threshold = simgrid::config::get_value<double>("smpi/cpu-threshold");
+ if (smpi_cpu_threshold < 0)
+ smpi_cpu_threshold = DBL_MAX;
+
+ smpi_host_speed = simgrid::config::get_value<double>("smpi/host-speed");
+ std::string smpi_privatize_option = simgrid::config::get_value<std::string>("smpi/privatization");
if (smpi_privatize_option == "no" || smpi_privatize_option == "0")
- smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE;
+ smpi_privatize_global_variables = SmpiPrivStrategies::NONE;
else if (smpi_privatize_option == "yes" || smpi_privatize_option == "1")
- smpi_privatize_global_variables = SMPI_PRIVATIZE_DEFAULT;
+ smpi_privatize_global_variables = SmpiPrivStrategies::DEFAULT;
else if (smpi_privatize_option == "mmap")
- smpi_privatize_global_variables = SMPI_PRIVATIZE_MMAP;
+ smpi_privatize_global_variables = SmpiPrivStrategies::MMAP;
else if (smpi_privatize_option == "dlopen")
- smpi_privatize_global_variables = SMPI_PRIVATIZE_DLOPEN;
+ smpi_privatize_global_variables = SmpiPrivStrategies::DLOPEN;
else
xbt_die("Invalid value for smpi/privatization: '%s'", smpi_privatize_option.c_str());
-#if defined(__FreeBSD__)
- if (smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) {
- XBT_INFO("mmap privatization is broken on FreeBSD, switching to dlopen privatization instead.");
- smpi_privatize_global_variables = SMPI_PRIVATIZE_DLOPEN;
- }
-#endif
-
- if (smpi_cpu_threshold < 0)
- smpi_cpu_threshold = DBL_MAX;
+ if (not SMPI_switch_data_segment) {
+ XBT_DEBUG("Running without smpi_main(); disable smpi/privatization.");
+ smpi_privatize_global_variables = SmpiPrivStrategies::NONE;
+ }
+ if (not HAVE_WORKING_MMAP && smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
+ XBT_INFO("mmap privatization is broken on this platform, switching to dlopen privatization instead.");
+ smpi_privatize_global_variables = SmpiPrivStrategies::DLOPEN;
+ }
- std::string val = xbt_cfg_get_string("smpi/shared-malloc");
- if ((val == "yes") || (val == "1") || (val == "on") || (val == "global")) {
- smpi_cfg_shared_malloc = shmalloc_global;
- } else if (val == "local") {
- smpi_cfg_shared_malloc = shmalloc_local;
- } else if ((val == "no") || (val == "0") || (val == "off")) {
- smpi_cfg_shared_malloc = shmalloc_none;
- } else {
- xbt_die("Invalid value '%s' for option smpi/shared-malloc. Possible values: 'on' or 'global', 'local', 'off'",
- val.c_str());
- }
+ std::string val = simgrid::config::get_value<std::string>("smpi/shared-malloc");
+ if ((val == "yes") || (val == "1") || (val == "on") || (val == "global")) {
+ smpi_cfg_shared_malloc = SharedMallocType::GLOBAL;
+ } else if (val == "local") {
+ smpi_cfg_shared_malloc = SharedMallocType::LOCAL;
+ } else if ((val == "no") || (val == "0") || (val == "off")) {
+ smpi_cfg_shared_malloc = SharedMallocType::NONE;
+ } else {
+ xbt_die("Invalid value '%s' for option smpi/shared-malloc. Possible values: 'on' or 'global', 'local', 'off'",
+ val.c_str());
+ }
}
typedef std::function<int(int argc, char *argv[])> smpi_entry_point_type;
typedef int (* smpi_c_entry_point_type)(int argc, char **argv);
typedef void (*smpi_fortran_entry_point_type)();
-static int smpi_run_entry_point(smpi_entry_point_type entry_point, std::vector<std::string> args)
+template <typename F>
+static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path, std::vector<std::string> args)
{
- char noarg[] = {'\0'};
- const int argc = args.size();
- std::unique_ptr<char*[]> argv(new char*[argc + 1]);
- for (int i = 0; i != argc; ++i)
- argv[i] = args[i].empty() ? noarg : &args[i].front();
- argv[argc] = nullptr;
-
- int res = entry_point(argc, argv.get());
+ // copy C strings, we need them writable
+ std::vector<char*>* 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()); });
+
+ // set argv[0] to executable_path
+ xbt_free((*args4argv)[0]);
+ (*args4argv)[0] = xbt_strdup(executable_path.c_str());
+
+#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);
+
+#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)
return 0;
}
+
// TODO, remove the number of functions involved here
static smpi_entry_point_type smpi_resolve_function(void* handle)
{
smpi_fortran_entry_point_type entry_point_fortran = (smpi_fortran_entry_point_type)dlsym(handle, "user_main_");
if (entry_point_fortran != nullptr) {
- return [entry_point_fortran](int argc, char** argv) {
- smpi_process_init(&argc, &argv);
+ return [entry_point_fortran](int, char**) {
entry_point_fortran();
return 0;
};
return smpi_entry_point_type();
}
-int smpi_main(const char* executable, int argc, char *argv[])
+static void smpi_copy_file(const std::string& src, const std::string& target, off_t fdin_size)
+{
+ 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());
+ int fdout = open(target.c_str(), O_CREAT | O_RDWR, S_IRWXU);
+ xbt_assert(fdout >= 0, "Cannot write into %s", target.c_str());
+
+ XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast<intmax_t>(fdin_size), target.c_str());
+#if SG_HAVE_SENDFILE
+ ssize_t sent_size = sendfile(fdout, fdin, NULL, fdin_size);
+ if (sent_size == fdin_size) {
+ 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<intmax_t>(fdin_size), errno, strerror(errno));
+ }
+#endif
+ // If this point is reached, sendfile() actually is not available. Copy file by hand.
+ const int bufsize = 1024 * 1024 * 4;
+ char* buf = new char[bufsize];
+ while (int got = read(fdin, buf, bufsize)) {
+ if (got == -1) {
+ xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str());
+ } else {
+ char* p = buf;
+ int todo = got;
+ while (int done = write(fdout, p, todo)) {
+ if (done == -1) {
+ xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str());
+ } else {
+ p += done;
+ todo -= done;
+ }
+ }
+ }
+ }
+ delete[] buf;
+ close(fdin);
+ close(fdout);
+}
+
+#if not defined(__APPLE__) && not defined(__HAIKU__)
+static int visit_libs(struct dl_phdr_info* info, size_t, void* data)
+{
+ char* libname = (char*)(data);
+ const char *path = info->dlpi_name;
+ if(strstr(path, libname)){
+ strncpy(libname, path, 512);
+ return 1;
+ }
+
+ return 0;
+}
+#endif
+
+static void smpi_init_privatization_dlopen(const std::string& executable)
+{
+ // Prepare the copy of the binary (get its size)
+ struct stat fdin_stat;
+ stat(executable.c_str(), &fdin_stat);
+ off_t fdin_size = fdin_stat.st_size;
+
+ std::string libnames = simgrid::config::get_value<std::string>("smpi/privatize-libs");
+ if (not libnames.empty()) {
+ // split option
+ std::vector<std::string> privatize_libs;
+ boost::split(privatize_libs, libnames, boost::is_any_of(";"));
+
+ 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);
+ // get library name from path
+ char fullpath[512] = {'\0'};
+ strncpy(fullpath, libname.c_str(), 511);
+#if not defined(__APPLE__) && not defined(__HAIKU__)
+ xbt_assert(0 != dl_iterate_phdr(visit_libs, fullpath),
+ "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath);
+ XBT_DEBUG("Extra lib to privatize '%s' found", fullpath);
+#else
+ xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku");
+#endif
+ privatize_libs_paths.push_back(fullpath);
+ dlclose(libhandle);
+ }
+ }
+
+ simix_global->default_function = [executable, fdin_size](std::vector<std::string> args) {
+ return std::function<void()>([executable, fdin_size, args] {
+ static std::size_t rank = 0;
+ // Copy the dynamic library:
+ std::string target_executable =
+ executable + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so";
+
+ smpi_copy_file(executable, target_executable, fdin_size);
+ // if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one.
+ std::vector<std::string> target_libs;
+ for (auto const& libpath : privatize_libs_paths) {
+ // if we were given a full path, strip it
+ size_t index = libpath.find_last_of("/\\");
+ std::string libname;
+ if (index != std::string::npos)
+ libname = libpath.substr(index + 1);
+
+ if (not libname.empty()) {
+ // load the library to add it to the local libs, to get the absolute path
+ struct stat fdin_stat2;
+ stat(libpath.c_str(), &fdin_stat2);
+ off_t fdin_size2 = fdin_stat2.st_size;
+
+ // 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.
+ unsigned int pad = 7;
+ if (libname.length() < pad)
+ pad = libname.length();
+ std::string target_lib =
+ std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
+ target_libs.push_back(target_lib);
+ XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2);
+ smpi_copy_file(libpath, target_lib, fdin_size2);
+
+ std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_lib + "/g' " + target_executable;
+ xbt_assert(system(sedcommand.c_str()) == 0, "error while applying sed command %s \n", sedcommand.c_str());
+ }
+ }
+
+ rank++;
+ // Load the copy and resolve the entry point:
+ void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND);
+ int saved_errno = errno;
+ if (simgrid::config::get_value<bool>("smpi/keep-temps") == false) {
+ unlink(target_executable.c_str());
+ for (const std::string& target_lib : target_libs)
+ unlink(target_lib.c_str());
+ }
+ xbt_assert(handle != nullptr, "dlopen failed: %s (errno: %d -- %s)", 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");
+ smpi_run_entry_point(entry_point, executable, args);
+ });
+ };
+}
+
+static void smpi_init_privatization_no_dlopen(const std::string& executable)
{
- srand(SMPI_RAND_SEED);
+ if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP)
+ smpi_prepare_global_memory_segment();
+
+ // Load the dynamic library and resolve the entry point:
+ void* handle = dlopen(executable.c_str(), RTLD_LAZY | RTLD_LOCAL);
+ xbt_assert(handle != nullptr, "dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno,
+ strerror(errno));
+ smpi_entry_point_type entry_point = smpi_resolve_function(handle);
+ xbt_assert(entry_point, "main not found in %s", executable.c_str());
+
+ if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP)
+ smpi_backup_global_memory_segment();
+
+ // Execute the same entry point for each simulated process:
+ simix_global->default_function = [entry_point, executable](std::vector<std::string> args) {
+ return std::function<void()>(
+ [entry_point, executable, args] { smpi_run_entry_point(entry_point, executable, args); });
+ };
+}
+int smpi_main(const char* executable, int argc, char* argv[])
+{
if (getenv("SMPI_PRETEND_CC") != nullptr) {
/* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
* configuration tools */
}
TRACE_global_init();
-
SIMIX_global_init(&argc, argv);
- MSG_init(&argc,argv);
SMPI_switch_data_segment = &smpi_switch_data_segment;
-
- // TODO This will not be executed in the case where smpi_main is not called,
- // e.g., not for smpi_msg_masterslave. This should be moved to another location
- // that is always called -- maybe close to Actor::onCreation?
- simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
- host.extension_set(new simgrid::smpi::SmpiHost(&host));
- });
-
+ sg_storage_file_system_init();
// parse the platform file: get the host list
- SIMIX_create_environment(argv[1]);
+ simgrid::s4u::Engine::get_instance()->load_platform(argv[1]);
SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback);
smpi_init_options();
- if (smpi_privatize_global_variables == SMPI_PRIVATIZE_DLOPEN) {
-
- std::string executable_copy = executable;
-
- // Prepare the copy of the binary (get its size)
- struct stat fdin_stat;
- stat(executable_copy.c_str(), &fdin_stat);
- off_t fdin_size = fdin_stat.st_size;
- static std::size_t rank = 0;
-
- simix_global->default_function = [executable_copy, fdin_size](std::vector<std::string> args) {
- return std::function<void()>([executable_copy, fdin_size, args] {
-
- // Copy the dynamic library:
- std::string target_executable = executable_copy
- + "_" + std::to_string(getpid())
- + "_" + std::to_string(rank++) + ".so";
-
- int fdin = open(executable_copy.c_str(), O_RDONLY);
- xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.",
- executable_copy.c_str());
- int fdout = open(target_executable.c_str(), O_CREAT | O_RDWR, S_IRWXU);
- xbt_assert(fdout >= 0, "Cannot write into %s", target_executable.c_str());
-
-#if HAVE_SENDFILE
- ssize_t sent_size = sendfile(fdout, fdin, NULL, fdin_size);
- xbt_assert(sent_size == fdin_size,
- "Error while copying %s: only %zd bytes copied instead of %ld (errno: %d -- %s)",
- target_executable.c_str(), sent_size, fdin_size, errno, strerror(errno));
-#else
- XBT_VERB("Copy %d bytes into %s", static_cast<int>(fdin_size), target_executable.c_str());
- const int bufsize = 1024 * 1024 * 4;
- char buf[bufsize];
- while (int got = read(fdin, buf, bufsize)) {
- if (got == -1) {
- xbt_assert(errno == EINTR, "Cannot read from %s", executable_copy.c_str());
- } else {
- char* p = buf;
- int todo = got;
- while (int done = write(fdout, p, todo)) {
- if (done == -1) {
- xbt_assert(errno == EINTR, "Cannot write into %s", target_executable.c_str());
- } else {
- p += done;
- todo -= done;
- }
- }
- }
- }
-#endif
- close(fdin);
- close(fdout);
-
- // Load the copy and resolve the entry point:
- void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);
- int saved_errno = errno;
- if (xbt_cfg_get_boolean("smpi/keep-temps") == false)
- unlink(target_executable.c_str());
- if (handle == nullptr)
- xbt_die("dlopen failed: %s (errno: %d -- %s)", dlerror(), saved_errno, strerror(saved_errno));
- smpi_entry_point_type entry_point = smpi_resolve_function(handle);
- if (not entry_point)
- xbt_die("Could not resolve entry point");
-
- smpi_run_entry_point(entry_point, args);
- });
- };
- }
- else {
- // Load the dynamic library and resolve the entry point:
- void* handle = dlopen(executable, RTLD_LAZY | RTLD_LOCAL);
- if (handle == nullptr)
- xbt_die("dlopen failed for %s: %s (errno: %d -- %s)", executable, dlerror(), errno, strerror(errno));
- smpi_entry_point_type entry_point = smpi_resolve_function(handle);
- if (not entry_point)
- xbt_die("main not found in %s", executable);
- // TODO, register the executable for SMPI privatization
-
- // Execute the same entry point for each simulated process:
- simix_global->default_function = [entry_point](std::vector<std::string> args) {
- return std::function<void()>([entry_point, args] {
- smpi_run_entry_point(entry_point, args);
- });
- };
- }
+ if (smpi_privatize_global_variables == SmpiPrivStrategies::DLOPEN)
+ smpi_init_privatization_dlopen(executable);
+ else
+ smpi_init_privatization_no_dlopen(executable);
SMPI_init();
- SIMIX_launch_application(argv[2]);
+ simgrid::s4u::Engine::get_instance()->load_deployment(argv[2]);
SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr,
process_data.size()); // This call has a side effect on process_count...
MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name);
- smpi_universe_size = process_count;
-
/* Clean IO before the run */
fflush(stdout);
SIMIX_run();
xbt_os_walltimer_stop(global_timer);
- if (xbt_cfg_get_boolean("smpi/display-timing")){
+ if (simgrid::config::get_value<bool>("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"
"You may want to use sampling functions or trace replay to reduce this.");
}
}
- smpi_global_destroy();
-
- TRACE_end();
+ SMPI_finalize();
return smpi_exit_status;
}
// Called either directly from the user code, or from the code called by smpirun
void SMPI_init(){
- simgrid::s4u::Actor::onCreation.connect([](simgrid::s4u::ActorPtr actor) {
- if (not actor->isDaemon()) {
- process_data.insert({actor, new simgrid::smpi::Process(actor, nullptr)});
+ simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::Actor& actor) {
+ if (not actor.is_daemon()) {
+ process_data.insert({&actor, new simgrid::smpi::ActorExt(&actor)});
}
});
- simgrid::s4u::Actor::onDestruction.connect([](simgrid::s4u::ActorPtr actor) {
- auto it = process_data.find(actor);
+ simgrid::s4u::Actor::on_destruction.connect([](simgrid::s4u::Actor const& actor) {
+ XBT_DEBUG("Delete the extension of actor %s", actor.get_cname());
+ auto it = process_data.find(&actor);
if (it != process_data.end()) {
delete it->second;
process_data.erase(it);
}
});
+ simgrid::s4u::Host::on_creation.connect(
+ [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); });
smpi_init_options();
- smpi_global_init();
+ if (not MC_is_active()) {
+ global_timer = xbt_os_timer_new();
+ xbt_os_walltimer_start(global_timer);
+ }
+
+ std::string filename = simgrid::config::get_value<std::string>("smpi/comp-adjustment-file");
+ if (not filename.empty()) {
+ std::ifstream fstream(filename);
+ xbt_assert(fstream.is_open(), "Could not open file %s. Does it exist?", filename.c_str());
+
+ std::string line;
+ typedef boost::tokenizer<boost::escaped_list_separator<char>> Tokenizer;
+ std::getline(fstream, line); // Skip the header line
+ while (std::getline(fstream, line)) {
+ Tokenizer tok(line);
+ Tokenizer::iterator it = tok.begin();
+ Tokenizer::iterator end = std::next(tok.begin());
+
+ std::string location = *it;
+ boost::trim(location);
+ location2speedup.insert(std::pair<std::string, double>(location, std::stod(*end)));
+ }
+ }
+ smpi_init_papi();
smpi_check_options();
- TRACE_smpi_alloc();
- simgrid::s4u::onSimulationEnd.connect(TRACE_smpi_release);
- if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
- smpi_backup_global_memory_segment();
}
-void SMPI_finalize(){
- smpi_global_destroy();
-}
+void SMPI_finalize()
+{
+ smpi_bench_destroy();
+ smpi_shared_destroy();
+ smpi_deployment_cleanup_instances();
-void smpi_mpi_init() {
- if(smpi_init_sleep > 0)
- simcall_process_sleep(smpi_init_sleep);
-}
+ if (simgrid::smpi::Colls::smpi_coll_cleanup_callback != nullptr)
+ simgrid::smpi::Colls::smpi_coll_cleanup_callback();
-double smpi_mpi_wtime(){
- double time;
- if (smpi_process()->initialized() != 0 && smpi_process()->finalized() == 0 && smpi_process()->sampling() == 0) {
- smpi_bench_end();
- time = SIMIX_get_clock();
- // to avoid deadlocks if used as a break condition, such as
- // while (MPI_Wtime(...) < time_limit) {
- // ....
- // }
- // because the time will not normally advance when only calls to MPI_Wtime
- // are made -> deadlock (MPI_Wtime never reaches the time limit)
- if(smpi_wtime_sleep > 0)
- simcall_process_sleep(smpi_wtime_sleep);
- smpi_bench_begin();
- } else {
- time = SIMIX_get_clock();
+ MPI_COMM_WORLD = MPI_COMM_NULL;
+
+ if (not MC_is_active()) {
+ xbt_os_timer_free(global_timer);
}
- return time;
+
+ if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP)
+ smpi_destroy_global_memory_segments();
+ if (simgrid::smpi::F2C::lookup() != nullptr)
+ simgrid::smpi::F2C::delete_lookup();
}
+void smpi_mpi_init() {
+ smpi_init_fortran_types();
+ if(smpi_init_sleep > 0)
+ simgrid::s4u::this_actor::sleep_for(smpi_init_sleep);
+}