/* 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 <spawn.h>
-#include <sys/types.h>
-#include <sys/wait.h>
-#include <dlfcn.h>
-
#include "mc/mc.h"
-#include "private.h"
-#include "private.hpp"
#include "simgrid/s4u/Mailbox.hpp"
-#include "smpi/smpi_shared_malloc.hpp"
-#include "simgrid/sg_config.h"
-#include "src/kernel/activity/SynchroComm.hpp"
-#include "src/mc/mc_record.h"
-#include "src/mc/mc_replay.h"
+#include "simgrid/s4u/Host.hpp"
#include "src/msg/msg_private.h"
#include "src/simix/smx_private.h"
-#include "surf/surf.h"
-#include "xbt/replay.hpp"
-#include <xbt/config.hpp>
+#include "src/surf/surf_interface.hpp"
+#include "src/smpi/SmpiHost.hpp"
+#include "xbt/config.hpp"
+#include "src/smpi/private.h"
+#include "src/smpi/smpi_coll.hpp"
+#include "src/smpi/smpi_comm.hpp"
+#include "src/smpi/smpi_group.hpp"
+#include "src/smpi/smpi_info.hpp"
+#include "src/smpi/smpi_process.hpp"
+#include <dlfcn.h>
+#include <fcntl.h>
+#include <sys/stat.h>
#include <float.h> /* DBL_MAX */
#include <fstream>
-#include <map>
-#include <stdint.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string>
-#include <utility>
-#include <vector>
-#include <memory>
+
+#if HAVE_SENDFILE
+#include <sys/sendfile.h>
+#endif
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)");
#include <boost/tokenizer.hpp>
#define RTLD_DEEPBIND 0
#endif
-/* Mac OSX does not have any header file providing that definition so we have to duplicate it here. Bummers. */
-extern char** environ; /* we use it in posix_spawnp below */
-
#if HAVE_PAPI
#include "papi.h"
const char* papi_default_config_name = "default";
void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t) = &smpi_comm_copy_buffer_callback;
-
-
int smpi_process_count()
{
return process_count;
simgrid::smpi::Process* smpi_process()
{
- simgrid::MsgActorExt* msgExt = static_cast<simgrid::MsgActorExt*>(SIMIX_process_self()->data);
+ smx_actor_t me = SIMIX_process_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->data);
return static_cast<simgrid::smpi::Process*>(msgExt->data);
}
return smpi_process()->index();
}
+void * smpi_process_get_user_data(){
+ return smpi_process()->get_user_data();
+}
+
+void smpi_process_set_user_data(void *data){
+ return smpi_process()->set_user_data(data);
+}
+
int smpi_global_size()
{
}
static void print(std::vector<std::pair<size_t, size_t>> vec) {
- fprintf(stderr, "{");
- for(auto elt: vec) {
- fprintf(stderr, "(0x%lx, 0x%lx),", elt.first, elt.second);
+ std::fprintf(stderr, "{");
+ for (auto elt : vec) {
+ std::fprintf(stderr, "(0x%zx, 0x%zx),", elt.first, elt.second);
}
- fprintf(stderr, "}\n");
+ std::fprintf(stderr, "}\n");
}
-static void memcpy_private(void *dest, const void *src, size_t n, std::vector<std::pair<size_t, size_t>> &private_blocks) {
+static void memcpy_private(void* dest, const void* src, std::vector<std::pair<size_t, size_t>>& private_blocks)
+{
for(auto block : private_blocks) {
memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first);
}
void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size)
{
- simgrid::kernel::activity::Comm *comm = dynamic_cast<simgrid::kernel::activity::Comm*>(synchro);
- int src_shared=0, dst_shared=0;
- size_t src_offset=0, dst_offset=0;
+ 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");
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::MsgActorExt*>(comm->src_proc->data)->data))->index()));
+ static_cast<simgrid::smpi::Process*>((static_cast<simgrid::msg::ActorExt*>(comm->src_proc->data)->data))
+ ->index());
tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
- memcpy_private(tmpbuff, buff, buff_size, private_blocks);
+ memcpy_private(tmpbuff, buff, private_blocks);
}
if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_start_data_exe)
&& ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){
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::MsgActorExt*>(comm->dst_proc->data)->data))->index()));
+ static_cast<simgrid::smpi::Process*>((static_cast<simgrid::msg::ActorExt*>(comm->dst_proc->data)->data))
+ ->index());
}
XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
- memcpy_private(comm->dst_buff, tmpbuff, buff_size, private_blocks);
+ memcpy_private(comm->dst_buff, tmpbuff, private_blocks);
if (comm->detached) {
// if this is a detached send, the source buffer was duplicated by SMPI
{
MPI_Group group;
- if (!MC_is_active()) {
+ if (not MC_is_active()) {
global_timer = xbt_os_timer_new();
xbt_os_walltimer_start(global_timer);
}
- if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') {
+ if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') {
std::string filename {xbt_cfg_get_string("smpi/comp-adjustment-file")};
std::ifstream fstream(filename);
- if (!fstream.is_open()) {
+ if (not fstream.is_open()) {
xbt_die("Could not open file %s. Does it exist?", filename.c_str());
}
MPI_COMM_WORLD = MPI_COMM_NULL;
- if (!MC_is_active()) {
+ if (not MC_is_active()) {
xbt_os_timer_free(global_timer);
}
}
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");
- const char* smpi_privatize_option = xbt_cfg_get_string("smpi/privatize-global-variables");
+ const char* smpi_privatize_option = xbt_cfg_get_string("smpi/privatization");
if (std::strcmp(smpi_privatize_option, "no") == 0)
smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE;
else if (std::strcmp(smpi_privatize_option, "yes") == 0)
smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE;
else
- xbt_die("Invalid value for smpi/privatize-global-variables: %s",
- smpi_privatize_option);
+ xbt_die("Invalid value for smpi/privatization: '%s'", smpi_privatize_option);
+
+#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;
+ }
+#endif
if (smpi_cpu_threshold < 0)
smpi_cpu_threshold = DBL_MAX;
char* val = xbt_cfg_get_string("smpi/shared-malloc");
- if (!strcasecmp(val, "yes") || !strcmp(val, "1") || !strcasecmp(val, "on") || !strcasecmp(val, "global")) {
+ if (not strcasecmp(val, "yes") || not strcmp(val, "1") || not strcasecmp(val, "on") ||
+ not strcasecmp(val, "global")) {
smpi_cfg_shared_malloc = shmalloc_global;
- } else if (!strcasecmp(val, "local")) {
+ } else if (not strcasecmp(val, "local")) {
smpi_cfg_shared_malloc = shmalloc_local;
- } else if (!strcasecmp(val, "no") || !strcmp(val, "0") || !strcasecmp(val, "off")) {
+ } else if (not strcasecmp(val, "no") || not strcmp(val, "0") || not strcasecmp(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'",
}
}
-static int execute_command(const char * const argv[])
-{
- pid_t pid;
- int status;
- if (posix_spawnp(&pid, argv[0], nullptr, nullptr, (char* const*) argv, environ) != 0)
- return 127;
- if (waitpid(pid, &status, 0) != pid)
- return 127;
- return status;
-}
-
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)(void);
+typedef void (*smpi_fortran_entry_point_type)();
static int smpi_run_entry_point(smpi_entry_point_type entry_point, 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() ? const_cast<char*>(""): &args[i].front();
+ argv[i] = args[i].empty() ? noarg : &args[i].front();
argv[argc] = nullptr;
int res = entry_point(argc, argv.get());
// TODO, remove the number of functions involved here
static smpi_entry_point_type smpi_resolve_function(void* handle)
{
- smpi_fortran_entry_point_type entry_point2 =
- (smpi_fortran_entry_point_type) dlsym(handle, "user_main_");
- if (entry_point2 != nullptr) {
- // fprintf(stderr, "EP user_main_=%p\n", entry_point2);
- return [entry_point2](int argc, char** argv) {
+ 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);
- entry_point2();
+ entry_point_fortran();
return 0;
};
}
- smpi_c_entry_point_type entry_point = (smpi_c_entry_point_type) dlsym(handle, "main");
+ smpi_c_entry_point_type entry_point = (smpi_c_entry_point_type)dlsym(handle, "main");
if (entry_point != nullptr) {
- // fprintf(stderr, "EP main=%p\n", entry_point);
return entry_point;
}
* configuration tools */
return 0;
}
- smpi_init_logs();
TRACE_global_init(&argc, argv);
- TRACE_add_start_function(TRACE_smpi_alloc);
- TRACE_add_end_function(TRACE_smpi_release);
SIMIX_global_init(&argc, argv);
MSG_init(&argc,argv);
SMPI_switch_data_segment = &smpi_switch_data_segment;
- smpi_init_options();
+ simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
+ host.extension_set(new simgrid::smpi::SmpiHost(&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);
- static std::size_t rank = 0;
+ smpi_init_options();
if (smpi_privatize_global_variables == SMPI_PRIVATIZE_DLOPEN) {
std::string executable_copy = executable;
- simix_global->default_function = [executable_copy](std::vector<std::string> args) {
- return std::function<void()>([executable_copy, args] {
+
+ // 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";
- // TODO, execute directly instead of relying on cp
- const char* command1 [] = {
- "cp", "--reflink=auto", "--", executable_copy.c_str(), target_executable.c_str(),
- nullptr
- };
- const char* command2 [] = {
- "cp", "--", executable_copy.c_str(), target_executable.c_str(),
- nullptr
- };
- if (execute_command(command1) != 0 && execute_command(command2) != 0)
- xbt_die("copy failed");
+
+ 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);
- unlink(target_executable.c_str());
+ if (xbt_cfg_get_boolean("smpi/keep-temps") == false)
+ unlink(target_executable.c_str());
if (handle == nullptr)
- xbt_die("dlopen failed");
+ xbt_die("dlopen failed: %s (errno: %d -- %s)", dlerror(), errno, strerror(errno));
smpi_entry_point_type entry_point = smpi_resolve_function(handle);
- if (!entry_point)
+ if (not entry_point)
xbt_die("Could not resolve entry point");
- smpi_run_entry_point(entry_point, args);
+ smpi_run_entry_point(entry_point, args);
});
};
// 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", executable);
+ 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 (!entry_point)
+ if (not entry_point)
xbt_die("main not found in %s", executable);
// TODO, register the executable for SMPI privatization
SIMIX_launch_application(argv[2]);
- smpi_global_init();
-
- smpi_check_options();
-
- if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
- smpi_initialize_global_memory_segments();
+ SMPI_init();
/* Clean IO before the run */
fflush(stdout);
if (MC_is_active()) {
MC_run();
} else {
-
+
SIMIX_run();
xbt_os_walltimer_stop(global_timer);
"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.");
}
}
int count = smpi_process_count();
- int i, ret=0;
- for (i = 0; i < count; i++) {
+ int ret = 0;
+ 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;
return ret;
}
-// This function can be called from extern file, to initialize logs, options, and processes of smpi
-// without the need of smpirun
+// Called either directly from the user code, or from the code called by smpirun
void SMPI_init(){
smpi_init_logs();
smpi_init_options();
smpi_global_init();
smpi_check_options();
- if (TRACE_is_enabled() && TRACE_is_configured())
- TRACE_smpi_alloc();
+ TRACE_smpi_alloc();
+ simgrid::surf::surfExitCallbacks.connect(TRACE_smpi_release);
if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
smpi_initialize_global_memory_segments();
}
}
void smpi_mpi_init() {
- if(smpi_init_sleep > 0)
+ if(smpi_init_sleep > 0)
simcall_process_sleep(smpi_init_sleep);
}
// }
// 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)
+ if(smpi_wtime_sleep > 0)
simcall_process_sleep(smpi_wtime_sleep);
smpi_bench_begin();
} else {