-/* Copyright (c) 2007-2017. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2007-2018. 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 "smpi_host.hpp"
#include "mc/mc.h"
-#include "private.hpp"
-#include "simgrid/s4u/Host.hpp"
-#include "simgrid/s4u/Mailbox.hpp"
+#include "simgrid/s4u/Engine.hpp"
#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 "src/simix/smx_private.hpp"
-#include "src/surf/surf_interface.hpp"
#include "xbt/config.hpp"
#include <cfloat> /* DBL_MAX */
#include <dlfcn.h>
#include <fcntl.h>
#include <fstream>
-#include <sys/stat.h>
+
+#if not defined(__APPLE__)
+#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
+# define HAVE_WORKING_MMAP (MAC_OS_X_VERSION_MIN_REQUIRED >= MAC_OS_X_VERSION_10_12)
+#elif defined(__FreeBSD__)
+# define HAVE_WORKING_MMAP 0
+#else
+# define HAVE_WORKING_MMAP 1
+#endif
#if HAVE_SENDFILE
#include <sys/sendfile.h>
#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;
-simgrid::smpi::Process **process_data = nullptr;
+static std::map</*process_id*/ ActorPtr, simgrid::smpi::Process*> process_data;
int process_count = 0;
+static int smpi_exit_status = 0;
int smpi_universe_size = 0;
-int* index_to_process_data = nullptr;
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
* "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 (basically, the one their SMPI instance uses).
+ * different communicators (the one their SMPI instance uses).
*
* See smpi_comm.cpp and the functions therein for details.
*/
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 char* smpi_default_instance_name = "smpirun";
+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(
simgrid::smpi::Process* smpi_process()
{
- smx_actor_t me = SIMIX_process_self();
+ 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->userdata);
+ simgrid::msg::ActorExt* msgExt = static_cast<simgrid::msg::ActorExt*>(me->get_impl()->get_user_data());
return static_cast<simgrid::smpi::Process*>(msgExt->data);
}
-simgrid::smpi::Process* smpi_process_remote(int index)
+simgrid::smpi::Process* smpi_process_remote(ActorPtr actor)
{
- return process_data[index_to_process_data[index]];
+ return process_data.at(actor);
}
MPI_Comm smpi_process_comm_self(){
}
int smpi_process_index(){
- return smpi_process()->index();
+ return simgrid::s4u::this_actor::get_pid();
}
void * smpi_process_get_user_data(){
smpi_comm_copy_data_callback = callback;
}
-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)
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(
- static_cast<simgrid::smpi::Process*>((static_cast<simgrid::msg::ActorExt*>(comm->src_proc->userdata)->data))
- ->index());
+ smpi_switch_data_segment(comm->src_proc->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) &&
+ 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(
- static_cast<simgrid::smpi::Process*>((static_cast<simgrid::msg::ActorExt*>(comm->dst_proc->userdata)->data))
- ->index());
+ smpi_switch_data_segment(comm->dst_proc->iface());
}
XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
memcpy_private(comm->dst_buff, tmpbuff, private_blocks);
/* nothing done in this version */
}
-static void smpi_check_options(){
+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"));
+ xbt_assert(simgrid::config::get_value<int>("smpi/async-small-thresh") <=
+ simgrid::config::get_value<int>("smpi/send-is-detached-thresh"));
- 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.");
- }
+ if (simgrid::config::is_default("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(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 process_data != nullptr;
+ return not process_data.empty();
}
void smpi_global_init()
xbt_os_walltimer_start(global_timer);
}
- std::string filename = xbt_cfg_get_string("smpi/comp-adjustment-file");
+ std::string filename = simgrid::config::get_value<std::string>("smpi/comp-adjustment-file");
if (not filename.empty()) {
std::ifstream fstream(filename);
if (not fstream.is_open()) {
// 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.
}
}
#endif
-
- if (index_to_process_data == nullptr) {
- index_to_process_data = new int[SIMIX_process_count()];
- }
-
- bool smpirun = 0;
- if (process_count == 0) { // The program has been dispatched but no other
- // SMPI instances have been registered. We're using smpirun.
- smpirun = true;
- SMPI_app_instance_register(smpi_default_instance_name, nullptr,
- SIMIX_process_count()); // 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;
- process_data = new simgrid::smpi::Process*[process_count];
- for (int i = 0; i < process_count; i++) {
- if (smpirun) {
- process_data[i] = new simgrid::smpi::Process(i, smpi_deployment_finalization_barrier(smpi_default_instance_name));
- smpi_deployment_register_process(smpi_default_instance_name, i, i);
- } else {
- // TODO We can pass a nullptr here because Process::set_data() assigns the
- // barrier from the instance anyway. This is ugly and should be changed
- process_data[i] = new simgrid::smpi::Process(i, nullptr);
- }
- }
}
void smpi_global_destroy()
smpi_bench_destroy();
smpi_shared_destroy();
smpi_deployment_cleanup_instances();
- int count = smpi_process_count();
- for (int i = 0; i < count; i++) {
- if(process_data[i]->comm_self()!=MPI_COMM_NULL){
- simgrid::smpi::Comm::destroy(process_data[i]->comm_self());
- }
- if(process_data[i]->comm_intra()!=MPI_COMM_NULL){
- simgrid::smpi::Comm::destroy(process_data[i]->comm_intra());
- }
- xbt_os_timer_free(process_data[i]->timer());
- xbt_mutex_destroy(process_data[i]->mailboxes_mutex());
- delete process_data[i];
- }
- delete[] process_data;
- process_data = nullptr;
if (simgrid::smpi::Colls::smpi_coll_cleanup_callback != nullptr)
simgrid::smpi::Colls::smpi_coll_cleanup_callback();
xbt_os_timer_free(global_timer);
}
- delete[] index_to_process_data;
- if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
+ if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP)
smpi_destroy_global_memory_segments();
smpi_free_static();
}
-extern "C" {
-
-static void smpi_init_logs(){
-
- /* Connect log categories. See xbt/log.c */
-
- XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this function: xbt_log_appender_file.c depends on it
- DO NOT connect this in XBT or so, or it will be useless to xbt_log_appender_file.c */
- XBT_LOG_CONNECT(instr_smpi);
- XBT_LOG_CONNECT(smpi_bench);
- XBT_LOG_CONNECT(smpi_coll);
- XBT_LOG_CONNECT(smpi_colls);
- XBT_LOG_CONNECT(smpi_comm);
- XBT_LOG_CONNECT(smpi_datatype);
- XBT_LOG_CONNECT(smpi_dvfs);
- XBT_LOG_CONNECT(smpi_group);
- XBT_LOG_CONNECT(smpi_host);
- XBT_LOG_CONNECT(smpi_kernel);
- XBT_LOG_CONNECT(smpi_mpi);
- XBT_LOG_CONNECT(smpi_memory);
- XBT_LOG_CONNECT(smpi_op);
- XBT_LOG_CONNECT(smpi_pmpi);
- XBT_LOG_CONNECT(smpi_process);
- XBT_LOG_CONNECT(smpi_request);
- XBT_LOG_CONNECT(smpi_replay);
- XBT_LOG_CONNECT(smpi_rma);
- XBT_LOG_CONNECT(smpi_shared);
- XBT_LOG_CONNECT(smpi_utils);
-}
-}
-
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");
+ smpi_host_speed = simgrid::config::get_value<double>("smpi/host-speed");
+ xbt_assert(smpi_host_speed > 0.0, "You're trying to set the host_speed to a non-positive value (given: %f)", 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("Mixing mmap privatization is broken on FreeBSD, switching to dlopen privatization instead.");
- smpi_privatize_global_variables = SMPI_PRIVATIZE_DLOPEN;
- }
+ if (not SMPI_switch_data_segment) {
+ XBT_DEBUG("Running without smpi_main(); disable smpi/privatization.");
+ smpi_privatize_global_variables = SmpiPrivStrategies::NONE;
+ }
+#if !HAVE_WORKING_MMAP
+ if (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;
+ }
#endif
- if (smpi_cpu_threshold < 0)
- smpi_cpu_threshold = DBL_MAX;
+ if (smpi_cpu_threshold < 0)
+ smpi_cpu_threshold = DBL_MAX;
- 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;
int res = entry_point(argc, argv.get());
if (res != 0){
XBT_WARN("SMPI process did not return 0. Return value : %d", res);
- smpi_process()->set_return_value(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)
{
return smpi_entry_point_type();
}
-int smpi_main(const char* executable, int argc, char *argv[])
+static void smpi_copy_file(std::string src, 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 %ld bytes into %s", static_cast<long>(fdin_size), target.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.c_str(), sent_size, fdin_size, errno, strerror(errno));
+#else
+ 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", 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;
+ }
+ }
+ }
+ }
+#endif
+ close(fdin);
+ close(fdout);
+}
+
+#if not defined(__APPLE__)
+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(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;
+ static std::size_t rank = 0;
+
+ 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'};
+ strcpy(fullpath, libname.c_str());
+#if not defined(__APPLE__)
+ int ret = dl_iterate_phdr(visit_libs, fullpath);
+ if (ret == 0)
+ xbt_die("Can't find a linked %s - check the setting you gave to smpi/privatize-libs", fullpath);
+ else
+ XBT_DEBUG("Extra lib to privatize found : %s", fullpath);
+#else
+ xbt_die("smpi/privatize-libs is not (yet) compatible with OSX");
+#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] {
+
+ // 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::string target_lib;
+ 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();
+ target_lib =
+ std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
+ 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;
+ int ret = system(sedcommand.c_str());
+ if (ret != 0)
+ xbt_die("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 | RTLD_DEEPBIND);
+ int saved_errno = errno;
+ if (simgrid::config::get_value<bool>("smpi/keep-temps") == false) {
+ unlink(target_executable.c_str());
+ if (not target_lib.empty())
+ unlink(target_lib.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);
+ });
+ };
+}
+
+static void smpi_init_privatization_no_dlopen(std::string executable)
+{
+ 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);
+ if (handle == nullptr)
+ xbt_die("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);
+ if (not entry_point)
+ xbt_die("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](std::vector<std::string> args) {
+ return std::function<void()>([entry_point, args] { smpi_run_entry_point(entry_point, args); });
+ };
+}
+
+int smpi_main(const char* executable, int argc, char* argv[])
{
srand(SMPI_RAND_SEED);
TRACE_global_init();
SIMIX_global_init(&argc, argv);
- MSG_init(&argc,argv);
+ MSG_init(&argc, argv); // FIXME Remove this MSG call. Once it's removed, we can remove the msg header include as well
SMPI_switch_data_segment = &smpi_switch_data_segment;
- simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
- host.extension_set(new simgrid::smpi::SmpiHost(&host));
- });
+ // 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::on_creation.connect(
+ [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); });
// parse the platform file: get the host list
SIMIX_create_environment(argv[1]);
SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback);
smpi_init_options();
+ if (smpi_privatize_global_variables == SmpiPrivStrategies::DLOPEN)
+ smpi_init_privatization_dlopen(executable);
+ else
+ smpi_init_privatization_no_dlopen(executable);
- 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", 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 | RTLD_DEEPBIND);
- 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);
- });
- };
-
- }
-
+ SMPI_init();
SIMIX_launch_application(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;
- SMPI_init();
/* 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.");
}
}
- int ret = 0;
- int count = smpi_process_count();
- for (int i = 0; i < count; i++) {
- if(process_data[i]->return_value()!=0){
- ret=process_data[i]->return_value();//return first non 0 value
- break;
- }
- }
smpi_global_destroy();
- TRACE_end();
-
- return ret;
+ return smpi_exit_status;
}
// Called either directly from the user code, or from the code called by smpirun
void SMPI_init(){
- smpi_init_logs();
+ simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::ActorPtr actor) {
+ if (not actor->is_daemon()) {
+ process_data.insert({actor, new simgrid::smpi::Process(actor, nullptr)});
+ }
+ });
+ simgrid::s4u::Actor::on_destruction.connect([](simgrid::s4u::ActorPtr actor) {
+ auto it = process_data.find(actor);
+ if (it != process_data.end()) {
+ delete it->second;
+ process_data.erase(it);
+ }
+ });
+
smpi_init_options();
smpi_global_init();
smpi_check_options();
- TRACE_smpi_alloc();
- simgrid::surf::surfExitCallbacks.connect(TRACE_smpi_release);
- if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
- smpi_backup_global_memory_segment();
+ simgrid::s4u::on_simulation_end.connect(TRACE_smpi_release);
}
void SMPI_finalize(){