Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
[mc] Disable timer in MC
[simgrid.git] / src / smpi / smpi_global.c
index 0072cda..4f001b6 100644 (file)
@@ -28,20 +28,22 @@ typedef struct s_smpi_process_data {
   smx_rdv_t mailbox_small;
   xbt_os_timer_t timer;
   MPI_Comm comm_self;
+  MPI_Comm comm_intra;
   MPI_Comm* comm_world;
   void *data;                   /* user data */
   int index;
   char state;
   int sampling;                 /* inside an SMPI_SAMPLE_ block? */
   char* instance_id;
+  int replaying;                /* is the process replaying a trace */
   xbt_bar_t finalization_barrier;
 } s_smpi_process_data_t;
 
 static smpi_process_data_t *process_data = NULL;
 int process_count = 0;
 int* index_to_process_data = NULL;
-
-
+extern double smpi_total_benched_time;
+xbt_os_timer_t global_timer;
 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
 int MPI_UNIVERSE_SIZE;
 
@@ -77,7 +79,7 @@ void smpi_process_init(int *argc, char ***argv)
     proc->context->cleanup_func=SIMIX_process_cleanup;
     char* instance_id = (*argv)[1];
     int rank = atoi((*argv)[2]);
-    index =  SIMIX_process_get_PID(proc) -1;
+    index = smpi_process_index_of_smx_process(proc);
 
     if(!index_to_process_data){
         index_to_process_data=(int*)xbt_malloc(SIMIX_process_count()*sizeof(int));
@@ -90,12 +92,12 @@ void smpi_process_init(int *argc, char ***argv)
     if(temp_bar != NULL) data->finalization_barrier = temp_bar;
     data->index = index;
     data->instance_id = instance_id;
+    data->replaying = 0;
     xbt_free(simcall_process_get_data(proc));
     simcall_process_set_data(proc, data);
     if (*argc > 3) {
       free((*argv)[1]);
-      free((*argv)[2]);
-      memmove(&(*argv)[1], &(*argv)[3], sizeof(char *) * (*argc - 3));
+      memmove(&(*argv)[0], &(*argv)[2], sizeof(char *) * (*argc - 2));
       (*argv)[(*argc) - 1] = NULL;
       (*argv)[(*argc) - 2] = NULL;
     }
@@ -104,10 +106,11 @@ void smpi_process_init(int *argc, char ***argv)
     data->argv = argv;
     // set the process attached to the mailbox
     simcall_rdv_set_receiver(data->mailbox_small, proc);
+
     XBT_DEBUG("<%d> New process in the game: %p", index, proc);
 
     if(smpi_privatize_global_variables){
-      switch_data_segment(index);
+      smpi_switch_data_segment(index);
     }
 
   }
@@ -119,7 +122,7 @@ void smpi_process_destroy(void)
 {
   int index = smpi_process_index();
   if(smpi_privatize_global_variables){
-    switch_data_segment(index);
+    smpi_switch_data_segment(index);
   }
   process_data[index_to_process_data[index]]->state = SMPI_FINALIZED;
   XBT_DEBUG("<%d> Process left the game", index);
@@ -130,10 +133,14 @@ void smpi_process_destroy(void)
  */
 void smpi_process_finalize(void)
 {
+    // This leads to an explosion of the search graph
+    // which cannot be reduced:
+    if(MC_is_active())
+      return;
+
     int index = smpi_process_index();
     // wait for all pending asynchronous comms to finish
     xbt_barrier_wait(process_data[index_to_process_data[index]]->finalization_barrier);
-
 }
 
 /**
@@ -164,10 +171,23 @@ int smpi_process_initialized(void)
 void smpi_process_mark_as_initialized(void)
 {
   int index = smpi_process_index();
-  if ((index != MPI_UNDEFINED) && (!process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
+  if ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
     process_data[index_to_process_data[index]]->state = SMPI_INITIALIZED;
 }
 
+void smpi_process_set_replaying(int value){
+  int index = smpi_process_index();
+  if ((index != MPI_UNDEFINED) && (process_data[index_to_process_data[index]]->state != SMPI_FINALIZED))
+    process_data[index_to_process_data[index]]->replaying = value;
+}
+
+int smpi_process_get_replaying(){
+  int index = smpi_process_index();
+  if (index != MPI_UNDEFINED)
+    return process_data[index_to_process_data[index]]->replaying;
+  else return _xbt_replay_is_active();
+}
+
 
 int smpi_global_size(void)
 {
@@ -277,6 +297,18 @@ MPI_Comm smpi_process_comm_self(void)
   return data->comm_self;
 }
 
+MPI_Comm smpi_process_get_comm_intra(void)
+{
+  smpi_process_data_t data = smpi_process_data();
+  return data->comm_intra;
+}
+
+void smpi_process_set_comm_intra(MPI_Comm comm)
+{
+  smpi_process_data_t data = smpi_process_data();
+  data->comm_intra = comm;
+}
+
 void smpi_process_set_sampling(int s)
 {
   smpi_process_data_t data = smpi_process_data();
@@ -289,19 +321,19 @@ int smpi_process_get_sampling(void)
   return data->sampling;
 }
 
+
 void print_request(const char *message, MPI_Request request)
 {
-  XBT_DEBUG
+  XBT_VERB
       ("%s  request %p  [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
        message, request, request->buf, request->size, request->src,
        request->dst, request->tag, request->flags);
 }
 
-void smpi_comm_copy_buffer_callback(smx_action_t comm,
+void smpi_comm_copy_buffer_callback(smx_synchro_t comm,
                                            void *buff, size_t buff_size)
 {
   XBT_DEBUG("Copy the data over");
-  if(_xbt_replay_is_active()) return;
   void* tmpbuff=buff;
 
   if((smpi_privatize_global_variables)
@@ -309,7 +341,7 @@ void smpi_comm_copy_buffer_callback(smx_action_t comm,
       && ((char*)buff < start_data_exe + size_data_exe )
     ){
        XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
-       switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
+       smpi_switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.src_proc))->index);
        tmpbuff = (void*)xbt_malloc(buff_size);
        memcpy(tmpbuff, buff, buff_size);
   }
@@ -320,7 +352,7 @@ void smpi_comm_copy_buffer_callback(smx_action_t comm,
       && ((char*)comm->comm.dst_buff < start_data_exe + size_data_exe )
     ){
        XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
-       switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
+       smpi_switch_data_segment(((smpi_process_data_t)SIMIX_process_get_data(comm->comm.dst_proc))->index);
   }
 
 
@@ -340,6 +372,13 @@ void smpi_comm_copy_buffer_callback(smx_action_t comm,
 
 }
 
+
+void smpi_comm_null_copy_buffer_callback(smx_synchro_t comm,
+                                           void *buff, size_t buff_size)
+{
+  return;
+}
+
 static void smpi_check_options(){
   //check correctness of MPI parameters
 
@@ -350,7 +389,7 @@ static void smpi_check_options(){
      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/running_power:<flops>\" to set its value."
-              "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information. ");
+              "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information.");
    }
 }
 
@@ -365,7 +404,10 @@ void smpi_global_init(void)
   char name[MAILBOX_NAME_MAXLEN];
   int smpirun=0;
 
-
+  if (!MC_is_active()) {
+    global_timer = xbt_os_timer_new();
+    xbt_os_walltimer_start(global_timer);
+  }
   if (process_count == 0){
     process_count = SIMIX_process_count();
     smpirun=1;
@@ -383,6 +425,7 @@ void smpi_global_init(void)
     if (MC_is_active())
       MC_ignore_heap(process_data[i]->timer, xbt_os_timer_size());
     process_data[i]->comm_self = MPI_COMM_NULL;
+    process_data[i]->comm_intra = MPI_COMM_NULL;
     process_data[i]->comm_world = NULL;
     process_data[i]->state = SMPI_UNINITIALIZED;
     process_data[i]->sampling = 0;
@@ -424,6 +467,10 @@ void smpi_global_destroy(void)
       smpi_group_unuse(smpi_comm_group(process_data[i]->comm_self));
       smpi_comm_destroy(process_data[i]->comm_self);
     }
+    if(process_data[i]->comm_intra!=MPI_COMM_NULL){
+      smpi_group_unuse(smpi_comm_group(process_data[i]->comm_intra));
+      smpi_comm_destroy(process_data[i]->comm_intra);
+    }
     xbt_os_timer_free(process_data[i]->timer);
     simcall_rdv_destroy(process_data[i]->mailbox);
     simcall_rdv_destroy(process_data[i]->mailbox_small);
@@ -481,6 +528,7 @@ static void smpi_init_logs(){
   XBT_LOG_CONNECT(smpi_mpi_dt);
   XBT_LOG_CONNECT(smpi_pmpi);
   XBT_LOG_CONNECT(smpi_replay);
+  XBT_LOG_CONNECT(smpi_rma);
 
 }
 
@@ -587,14 +635,14 @@ int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
 
   SIMIX_global_init(&argc, argv);
 
+  smpi_init_options();
+
   // parse the platform file: get the host list
   SIMIX_create_environment(argv[1]);
   SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_buffer_callback);
   SIMIX_function_register_default(realmain);
   SIMIX_launch_application(argv[2]);
 
-  smpi_init_options();
-
   smpi_global_init();
 
   smpi_check_options();
@@ -606,13 +654,25 @@ int smpi_main(int (*realmain) (int argc, char *argv[]), int argc, char *argv[])
   fflush(stdout);
   fflush(stderr);
 
-  if (MC_is_active())
+  if (MC_is_active()) {
     MC_do_the_modelcheck_for_real();
-  else
+  } else {
+  
     SIMIX_run();
 
-  if (sg_cfg_get_boolean("smpi/display_timing"))
-    XBT_INFO("Simulation time: %g seconds.", SIMIX_get_clock());
+    xbt_os_walltimer_stop(global_timer);
+    if (sg_cfg_get_boolean("smpi/display_timing")){
+      double global_time = xbt_os_timer_elapsed(global_timer);
+      XBT_INFO("Simulated time: %g seconds. \n "
+          "The simulation took %g seconds (after parsing and platform setup)\n"
+          "%g seconds were actual computation of the application"
+          , SIMIX_get_clock(), global_time , smpi_total_benched_time);
+          
+      if (smpi_total_benched_time/global_time>=0.75)
+      XBT_INFO("More than 75%% of the time was spent inside the application code.\n"
+      "You may want to use sampling functions or trace replay to reduce this.");
+    }
+  }
 
   smpi_global_destroy();