set(HAVE_THREAD_LOCAL_STORAGE 0)
endif()
+CHECK_INCLUDE_FILE("sys/sendfile.h" HAVE_SENDFILE_H)
+CHECK_FUNCTION_EXISTS(sendfile HAVE_SENDFILE)
+if(HAVE_SENDFILE_H AND HAVE_SENDFILE)
+ set(HAVE_SENDFILE 1)
+else()
+ set(HAVE_SENDFILE 0)
+endif()
+
if(enable_model-checking AND NOT "${CMAKE_SYSTEM}" MATCHES "Linux|FreeBSD")
message(WARNING "Support for model-checking has not been enabled on ${CMAKE_SYSTEM}: disabling it")
set(enable_model-checking FALSE)
p Testing the Chord implementation with MSG
+! timeout 30
! output sort 19
$ $SG_TEST_EXENV ${bindir:=.}/dht-chord$EXEEXT -nb_bits=6 ${srcdir:=.}/cluster.xml ${srcdir:=.}/../msg/dht-chord/dht-chord_d.xml --log=msg_chord.thres:verbose "--log=root.fmt:[%10.5r]%e(%i:%P@%h)%e%m%n"
> [ 0.00000] (1:node@node-0.acme.org) My finger table:
XBT_PUBLIC(void) TRACE_help(int detailed);
XBT_PUBLIC(void) TRACE_surf_resource_utilization_alloc();
XBT_PUBLIC(void) TRACE_surf_resource_utilization_release();
-XBT_PUBLIC(void) TRACE_add_start_function(void (*func)(void));
-XBT_PUBLIC(void) TRACE_add_end_function(void (*func)(void));
SG_END_DECL()
-#endif
\ No newline at end of file
+#endif
trace_precision = xbt_cfg_get_int(OPT_TRACING_PRECISION);
}
-static std::vector<std::function<void()>> TRACE_start_functions;
-
-void TRACE_add_start_function(void (*func) ())
-{
- TRACE_start_functions.push_back(func);
-}
-
int TRACE_start()
{
if (TRACE_is_configured())
user_host_variables = xbt_dict_new_homogeneous(xbt_free_f);
user_vm_variables = xbt_dict_new_homogeneous(xbt_free_f);
user_link_variables = xbt_dict_new_homogeneous(xbt_free_f);
-
- for (auto func: TRACE_start_functions)
- func();
}
- TRACE_start_functions.clear();
return 0;
}
-static std::vector<std::function<void()>> TRACE_end_functions;
-void TRACE_add_end_function(void (*func) (void))
-{
- TRACE_end_functions.push_back(func);
-}
-
int TRACE_end()
{
int retval;
PJ_container_release();
PJ_type_release();
- for (auto func: TRACE_end_functions)
- func();
- TRACE_start_functions.clear();
-
xbt_dict_free(&user_link_variables);
xbt_dict_free(&user_host_variables);
xbt_dict_free(&user_vm_variables);
--- /dev/null
+#include "smpi/smpi_utils.hpp"
+#include "src/smpi/SmpiHost.hpp"
+#include <simgrid/s4u/VirtualMachine.hpp>
+#include <string>
+#include <vector>
+
+XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_host, smpi, "Logging specific to SMPI (host)");
+
+namespace simgrid {
+namespace smpi {
+
+simgrid::xbt::Extension<simgrid::s4u::Host, SmpiHost> SmpiHost::EXTENSION_ID;
+
+double SmpiHost::orecv(size_t size)
+{
+ double current = orecv_parsed_values.empty() ? 0.0 : orecv_parsed_values.front().values[0] +
+ orecv_parsed_values.front().values[1] * size;
+
+ // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
+ // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
+ // Note: parse_factor() (used before) already sorts the vector we iterate over!
+ for (auto fact : orecv_parsed_values) {
+ if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
+ XBT_DEBUG("or : %zu <= %zu return %.10f", size, fact.factor, current);
+ return current;
+ } else {
+ // If the next section is too large, the current section must be used.
+ // Hence, save the cost, as we might have to use it.
+ current=fact.values[0]+fact.values[1]*size;
+ }
+ }
+ XBT_DEBUG("smpi_or: %zu is larger than largest boundary, return %.10f", size, current);
+
+ return current;
+}
+
+double SmpiHost::osend(size_t size)
+{
+ double current =
+ osend_parsed_values.empty() ? 0.0 : osend_parsed_values[0].values[0] + osend_parsed_values[0].values[1] * size;
+ // Iterate over all the sections that were specified and find the right
+ // value. (fact.factor represents the interval sizes; we want to find the
+ // section that has fact.factor <= size and no other such fact.factor <= size)
+ // Note: parse_factor() (used before) already sorts the vector we iterate over!
+ for (auto& fact : osend_parsed_values) {
+ if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
+ XBT_DEBUG("os : %zu <= %zu return %.10f", size, fact.factor, current);
+ return current;
+ } else {
+ // If the next section is too large, the current section must be used.
+ // Hence, save the cost, as we might have to use it.
+ current = fact.values[0] + fact.values[1] * size;
+ }
+ }
+ XBT_DEBUG("Searching for smpi/os: %zu is larger than the largest boundary, return %.10f", size, current);
+
+ return current;
+}
+
+double SmpiHost::oisend(size_t size)
+{
+ double current =
+ oisend_parsed_values.empty() ? 0.0 : oisend_parsed_values[0].values[0] + oisend_parsed_values[0].values[1] * size;
+
+ // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
+ // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
+ // Note: parse_factor() (used before) already sorts the vector we iterate over!
+ for (auto& fact : oisend_parsed_values) {
+ if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
+ XBT_DEBUG("ois : %zu <= %zu return %.10f", size, fact.factor, current);
+ return current;
+ } else {
+ // If the next section is too large, the current section must be used.
+ // Hence, save the cost, as we might have to use it.
+ current = fact.values[0] + fact.values[1] * size;
+ }
+ }
+ XBT_DEBUG("Searching for smpi/ois: %zu is larger than the largest boundary, return %.10f", size, current);
+
+ return current;
+}
+
+SmpiHost::SmpiHost(simgrid::s4u::Host *ptr) : host(ptr)
+{
+ if (!SmpiHost::EXTENSION_ID.valid())
+ SmpiHost::EXTENSION_ID = simgrid::s4u::Host::extension_create<SmpiHost>();
+
+ const char* orecv_string = host->property("smpi/or");
+ if (orecv_string != nullptr) {
+ orecv_parsed_values = parse_factor(orecv_string);
+ } else {
+ orecv_parsed_values = parse_factor(xbt_cfg_get_string("smpi/or"));
+ }
+
+ const char* osend_string = host->property("smpi/os");
+ if (osend_string != nullptr) {
+ osend_parsed_values = parse_factor(osend_string);
+ } else {
+ osend_parsed_values = parse_factor(xbt_cfg_get_string("smpi/os"));
+ }
+
+ const char* oisend_string = host->property("smpi/ois");
+ if (oisend_string != nullptr) {
+ oisend_parsed_values = parse_factor(oisend_string);
+ } else {
+ oisend_parsed_values = parse_factor(xbt_cfg_get_string("smpi/ois"));
+ }
+}
+
+SmpiHost::~SmpiHost()=default;
+
+static void onCreation(simgrid::s4u::Host& host)
+{
+}
+
+static void onHostDestruction(simgrid::s4u::Host& host)
+{
+ // Ignore virtual machines
+ if (dynamic_cast<simgrid::s4u::VirtualMachine*>(&host))
+ return;
+}
+
+void sg_smpi_host_init()
+{
+ simgrid::s4u::Host::onCreation.connect(&onCreation);
+ simgrid::s4u::Host::onDestruction.connect(&onHostDestruction);
+}
+
+}
+}
--- /dev/null
+#ifndef SMPI_HOST_HPP_
+#define SMPI_HOST_HPP_
+
+#include "src/include/smpi/smpi_utils.hpp"
+
+#include <simgrid/s4u/host.hpp>
+#include <string>
+#include <vector>
+#include <xbt/config.hpp>
+#include <xbt/Extendable.hpp>
+
+
+
+namespace simgrid {
+namespace smpi {
+
+void sg_smpi_host_init();
+static void onHostDestruction(simgrid::s4u::Host& host);
+static void onCreation(simgrid::s4u::Host& host);
+
+class SmpiHost {
+
+ private:
+ std::vector<s_smpi_factor_t> orecv_parsed_values;
+ std::vector<s_smpi_factor_t> osend_parsed_values;
+ std::vector<s_smpi_factor_t> oisend_parsed_values;
+ simgrid::s4u::Host *host = nullptr;
+
+ public:
+ static simgrid::xbt::Extension<simgrid::s4u::Host, SmpiHost> EXTENSION_ID;
+
+ explicit SmpiHost(simgrid::s4u::Host *ptr);
+ ~SmpiHost();
+
+ double orecv(size_t size);
+ double osend(size_t size);
+ double oisend(size_t size);
+};
+
+}
+}
+#endif
#include "private.h"
#include "simgrid/msg.h" /* barrier */
+#include "src/smpi/SmpiHost.hpp"
#include "xbt/dict.h"
#include "xbt/log.h"
#include "xbt/sysdep.h"
s_smpi_mpi_instance_t* instance = (s_smpi_mpi_instance_t*)xbt_malloc(sizeof(s_smpi_mpi_instance_t));
+ static int already_called = 0;
+ if (!already_called) {
+ already_called = 1;
+ xbt_dynar_t hosts = MSG_hosts_as_dynar();
+ unsigned int cursor;
+ void* h;
+ xbt_dynar_foreach(hosts, cursor, h) {
+ simgrid::s4u::Host* host = static_cast<simgrid::s4u::Host*>(h);
+ host->extension_set(new simgrid::smpi::SmpiHost(host));
+ }
+ xbt_dynar_free(&hosts);
+ }
+
instance->name = name;
instance->size = num_processes;
instance->present_processes = 0;
smpi_instances = xbt_dict_new_homogeneous(xbt_free_f);
}
+
xbt_dict_set(smpi_instances, name, (void*)instance, nullptr);
}
/* 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 <dlfcn.h>
+#include <fcntl.h>
#include <spawn.h>
+#include <sys/stat.h>
#include <sys/types.h>
#include <sys/wait.h>
-#include <dlfcn.h>
#include "mc/mc.h"
#include "private.h"
#include "src/mc/mc_replay.h"
#include "src/msg/msg_private.h"
#include "src/simix/smx_private.h"
+#include "src/surf/surf_interface.hpp"
+#include "src/smpi/SmpiHost.hpp"
#include "surf/surf.h"
#include "xbt/replay.hpp"
#include <xbt/config.hpp>
#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>
#include <boost/algorithm/string.hpp> /* trim_right / trim_left */
* 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]);
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 (open the file and get its size)
+ // (fdin will remain open for the whole process execution. That's a sort of leak but we can live with it)
+ int fdin = open(executable_copy.c_str(), O_RDONLY);
+ xbt_assert(fdin >= 0, "Cannot read from %s", executable_copy.c_str());
+ struct stat fdin_stat;
+ fstat(fdin, &fdin_stat);
+ off_t fdin_size = fdin_stat.st_size;
+
+ simix_global->default_function = [executable_copy, fdin, fdin_size](std::vector<std::string> args) {
+ return std::function<void()>([executable_copy, fdin, 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 fdout = open(target_executable.c_str(), O_WRONLY);
+ xbt_assert(fdout >= 0, "Cannot write into %s", target_executable.c_str());
+
+#if HAVE_SENDFILE
+ sendfile(fdout, fdin, NULL, fdin_size);
+#else
+ XBT_WARN("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(fdout);
// Load the copy and resolve the entry point:
void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);
if (!entry_point)
xbt_die("Could not resolve entry point");
- smpi_run_entry_point(entry_point, args);
+ smpi_run_entry_point(entry_point, args);
});
};
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);
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();
}
#include "src/simix/smx_private.h"
#include "simgrid/sg_config.h"
#include "smpi/smpi_utils.hpp"
+#include "src/smpi/SmpiHost.hpp"
#include <simgrid/s4u/host.hpp>
#include "src/kernel/activity/SynchroComm.hpp"
static simgrid::config::Flag<double> smpi_test_sleep(
"smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
-std::vector<s_smpi_factor_t> smpi_os_values;
-std::vector<s_smpi_factor_t> smpi_or_values;
std::vector<s_smpi_factor_t> smpi_ois_values;
extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
-static double smpi_os(size_t size)
-{
- if (smpi_os_values.empty()) {
- smpi_os_values = parse_factor(xbt_cfg_get_string("smpi/os"));
- }
- double current=smpi_os_values.empty()?0.0:smpi_os_values[0].values[0]+smpi_os_values[0].values[1]*size;
- // Iterate over all the sections that were specified and find the right
- // value. (fact.factor represents the interval sizes; we want to find the
- // section that has fact.factor <= size and no other such fact.factor <= size)
- // Note: parse_factor() (used before) already sorts the vector we iterate over!
- for (auto& fact : smpi_os_values) {
- if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
- XBT_DEBUG("os : %zu <= %zu return %.10f", size, fact.factor, current);
- return current;
- }else{
- // If the next section is too large, the current section must be used.
- // Hence, save the cost, as we might have to use it.
- current = fact.values[0]+fact.values[1]*size;
- }
- }
- XBT_DEBUG("Searching for smpi/os: %zu is larger than the largest boundary, return %.10f", size, current);
-
- return current;
-}
-
-static double smpi_ois(size_t size)
-{
- if (smpi_ois_values.empty()) {
- smpi_ois_values = parse_factor(xbt_cfg_get_string("smpi/ois"));
- }
- double current=smpi_ois_values.empty()?0.0:smpi_ois_values[0].values[0]+smpi_ois_values[0].values[1]*size;
- // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
- // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
- // Note: parse_factor() (used before) already sorts the vector we iterate over!
- for (auto& fact : smpi_ois_values) {
- if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
- XBT_DEBUG("ois : %zu <= %zu return %.10f", size, fact.factor, current);
- return current;
- }else{
- // If the next section is too large, the current section must be used.
- // Hence, save the cost, as we might have to use it.
- current = fact.values[0]+fact.values[1]*size;
- }
- }
- XBT_DEBUG("Searching for smpi/ois: %zu is larger than the largest boundary, return %.10f", size, current);
-
- return current;
-}
-
-static double smpi_or(size_t size)
-{
- if (smpi_or_values.empty()) {
- smpi_or_values = parse_factor(xbt_cfg_get_string("smpi/or"));
- }
-
- double current=smpi_or_values.empty()?0.0:smpi_or_values.front().values[0]+smpi_or_values.front().values[1]*size;
-
- // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
- // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
- // Note: parse_factor() (used before) already sorts the vector we iterate over!
- for (auto fact : smpi_or_values) {
- if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
- XBT_DEBUG("or : %zu <= %zu return %.10f", size, fact.factor, current);
- return current;
- } else {
- // If the next section is too large, the current section must be used.
- // Hence, save the cost, as we might have to use it.
- current=fact.values[0]+fact.values[1]*size;
- }
- }
- XBT_DEBUG("smpi_or: %zu is larger than largest boundary, return %.10f", size, current);
-
- return current;
-}
-
namespace simgrid{
namespace smpi{
if(!(old_type_->flags() & DT_FLAG_DERIVED)){
oldbuf = buf_;
if (!process->replaying() && oldbuf != nullptr && size_!=0){
- if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
+ if((smpi_privatize_global_variables != 0)
&& (static_cast<char*>(buf_) >= smpi_start_data_exe)
&& (static_cast<char*>(buf_) < smpi_start_data_exe + smpi_size_data_exe )){
XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
//if we are giving back the control to the user without waiting for completion, we have to inject timings
double sleeptime = 0.0;
- if(detached_ != 0 || ((flags_ & (ISEND|SSEND)) != 0)){// issend should be treated as isend
- //isend and send timings may be different
- sleeptime = ((flags_ & ISEND) != 0) ? smpi_ois(size_) : smpi_os(size_);
+ if (detached_ != 0 || ((flags_ & (ISEND | SSEND)) != 0)) { // issend should be treated as isend
+ // isend and send timings may be different
+ sleeptime = ((flags_ & ISEND) != 0)
+ ? simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->oisend(size_)
+ : simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->osend(size_);
}
if(sleeptime > 0.0){
if((((req->flags_ & ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED))){// && (!smpi_is_shared(req->old_buf_))){
if (!smpi_process()->replaying()){
- if( smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
+ if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
&& ((char*)req->old_buf_ < smpi_start_data_exe + smpi_size_data_exe )){
XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
smpi_switch_data_segment(smpi_process()->index());
}
if(req->detached_sender_ != nullptr){
//integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
- double sleeptime = smpi_or(req->real_size_);
+ double sleeptime = simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->orecv(req->real_size());
if(sleeptime > 0.0){
simcall_process_sleep(sleeptime);
XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
if (!future_evt_set)
future_evt_set = new simgrid::trace_mgr::future_evt_set();
- TRACE_add_start_function(TRACE_surf_alloc);
- TRACE_add_end_function(TRACE_surf_release);
+ TRACE_surf_alloc();
+ simgrid::surf::surfExitCallbacks.connect(TRACE_surf_release);
sg_config_init(argc, argv);
src/smpi/smpi_f2c.hpp
src/smpi/smpi_group.cpp
src/smpi/smpi_group.hpp
+ src/smpi/SmpiHost.cpp
+ src/smpi/SmpiHost.hpp
src/smpi/smpi_mpi.cpp
src/smpi/smpi_datatype.cpp
src/smpi/smpi_datatype.hpp
#cmakedefine01 HAVE_PRIVATIZATION
/* We have PAPI to fine-grain trace execution time */
#cmakedefine01 HAVE_PAPI
+/* We have sendfile to efficiently copy files for dl-open privatization */
+#cmakedefine01 HAVE_SENDFILE
/* Other function checks */
/* Function backtrace */
