Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
use sensible process names (their rank) in SMPI, and get simix ignore that
[simgrid.git] / src / smpi / smpi_global.c
index 64a4b31..6fa73d2 100644 (file)
+/* Copyright (c) 2007, 2008, 2009, 2010. 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 <stdint.h>
 #include <stdio.h>
+#include <stdlib.h>
 
 #include "private.h"
+#include "smpi_mpi_dt_private.h"
+#include "mc/mc.h"
+#include "surf/surf.h"
 
-XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi, XBT_LOG_ROOT_CAT, "All SMPI categories");
+XBT_LOG_NEW_CATEGORY(smpi, "All SMPI categories");
 
-SMPI_Global_t     smpi_global     = NULL;
+XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi,
+                                "Logging specific to SMPI (kernel)");
 
-void *smpi_request_new()
-{
-       smpi_mpi_request_t request = xbt_new(s_smpi_mpi_request_t, 1);
+typedef struct s_smpi_process_data {
+  int index;
+  int* argc;
+  char*** argv;
+  smx_rdv_t mailbox;
+  xbt_os_timer_t timer;
+  double simulated;
+  MPI_Comm comm_self;
+} s_smpi_process_data_t;
+
+static smpi_process_data_t *process_data = NULL;
+static int process_count = 0;
+
+MPI_Comm MPI_COMM_WORLD = MPI_COMM_NULL;
 
-       request->completed = 0;
-       request->mutex     = SIMIX_mutex_init();
-       request->cond      = SIMIX_cond_init();
+#define MAILBOX_NAME_MAXLEN (5 + sizeof(int) * 2 + 1)
 
-       return request;
+static char* get_mailbox_name(char* str, int index) {
+  snprintf(str, MAILBOX_NAME_MAXLEN, "SMPI-%0*x", (int)(sizeof(int) * 2), index);
+  return str;
 }
 
-void smpi_request_free(void *pointer)
+void smpi_process_init(int *argc, char ***argv)
 {
+  int index;
+  smpi_process_data_t data;
+  smx_process_t proc;
+
+  if(argc && argv) {
+    proc = SIMIX_process_self();
+    index = atoi((*argv)[1]);
+    data = smpi_process_remote_data(index);
+    SIMIX_req_process_set_data(proc, data);
+    if (*argc > 2) {
+      free((*argv)[1]);
+      memmove(&(*argv)[1], &(*argv)[2], sizeof(char *) * (*argc - 2));
+      (*argv)[(*argc) - 1] = NULL;
+    }
+    (*argc)--;
+    data->argc = argc;
+    data->argv = argv;
+    XBT_DEBUG("<%d> New process in the game: %p", index, proc);
+  }
+}
+
+void smpi_process_destroy(void)
+{
+  int index = smpi_process_index();
 
-       smpi_mpi_request_t request = pointer;
+  XBT_DEBUG("<%d> Process left the game", index);
+}
 
-       if (NULL != request) {
-               SIMIX_cond_destroy(request->cond);
-               SIMIX_mutex_destroy(request->mutex);
-               xbt_free(request);
-       }
+int smpi_process_argc(void) {
+  smpi_process_data_t data = smpi_process_data();
 
-       return;
+  return data->argc ? *(data->argc) - 1 : 0;
 }
 
-void smpi_request_reset(void *pointer)
-{
-       return;
+int smpi_process_getarg(integer* index, char* dst, ftnlen len) {
+  smpi_process_data_t data = smpi_process_data();
+  char* arg;
+  ftnlen i;
+
+  if(!data->argc || !data->argv
+     || *index < 1 || *index >= *(data->argc)) {
+    return -1;
+  }
+  arg = (*data->argv)[*index];
+  for(i = 0; i < len && arg[i] != '\0'; i++) {
+    dst[i] = arg[i];
+  }
+  for(; i < len; i++) {
+    dst[i] = ' ';
+  }
+  return 0;
 }
 
+int smpi_global_size(void) {
+   char* value = getenv("SMPI_GLOBAL_SIZE");
 
-void *smpi_message_new()
-{
-       return xbt_new(s_smpi_received_message_t, 1);
+   if(!value) {
+      fprintf(stderr, "Please set env var SMPI_GLOBAL_SIZE to expected number of processes.\n");
+      abort();
+   }
+   return atoi(value);
 }
 
-void smpi_message_free(void *pointer)
+smpi_process_data_t smpi_process_data(void)
 {
-       if (NULL != pointer) {
-               xbt_free(pointer);
-       }
-
-       return;
+  return SIMIX_process_self_get_data(SIMIX_process_self());
 }
 
-void smpi_message_reset(void *pointer)
+smpi_process_data_t smpi_process_remote_data(int index)
 {
-       return;
+  return process_data[index];
 }
 
-void smpi_global_init()
+int smpi_process_count(void)
 {
-       int i;
-
-       int size = SIMIX_host_get_number();
-
-       smpi_global                                      = xbt_new(s_SMPI_Global_t, 1);
+  return process_count;
+}
 
-       // config variable
-       smpi_global->reference_speed                     = SMPI_DEFAULT_SPEED;
+int smpi_process_index(void)
+{
+  smpi_process_data_t data = smpi_process_data();
 
-       smpi_global->root_ready                          = 0;
-       smpi_global->ready_process_count                 = 0;
+  return data->index;
+}
 
-       // start/stop
-       smpi_global->start_stop_mutex                    = SIMIX_mutex_init();
-       smpi_global->start_stop_cond                     = SIMIX_cond_init();
+smx_rdv_t smpi_process_mailbox(void) {
+  smpi_process_data_t data = smpi_process_data();
 
-       // processes
-       smpi_global->sender_processes                    = xbt_new(smx_process_t, size);
-       smpi_global->receiver_processes                  = xbt_new(smx_process_t, size);
+  return data->mailbox;
+}
 
-       // running hosts
-       smpi_global->running_hosts_count_mutex           = SIMIX_mutex_init();
-        smpi_global->running_hosts_count                 = 0;
+smx_rdv_t smpi_process_remote_mailbox(int index) {
+  smpi_process_data_t data = smpi_process_remote_data(index);
 
-       // mallocators
-       smpi_global->request_mallocator                  = xbt_mallocator_new(SMPI_REQUEST_MALLOCATOR_SIZE,
-                                                            smpi_request_new, smpi_request_free, smpi_request_reset);
-       smpi_global->message_mallocator                  = xbt_mallocator_new(SMPI_MESSAGE_MALLOCATOR_SIZE,
-                                                            smpi_message_new, smpi_message_free, smpi_message_reset);
+  return data->mailbox;
+}
 
-       // queues
-       smpi_global->pending_send_request_queues         = xbt_new(xbt_fifo_t,  size);
-       smpi_global->pending_send_request_queues_mutexes = xbt_new(smx_mutex_t, size);
-       smpi_global->pending_recv_request_queues         = xbt_new(xbt_fifo_t,  size);
-       smpi_global->pending_recv_request_queues_mutexes = xbt_new(smx_mutex_t, size);
-       smpi_global->received_message_queues             = xbt_new(xbt_fifo_t,  size);
-       smpi_global->received_message_queues_mutexes     = xbt_new(smx_mutex_t, size);
+xbt_os_timer_t smpi_process_timer(void)
+{
+  smpi_process_data_t data = smpi_process_data();
 
-       // timers
-       smpi_global->timers                              = xbt_new(xbt_os_timer_t, size);
-       smpi_global->timers_mutexes                      = xbt_new(smx_mutex_t, size);
+  return data->timer;
+}
 
-       for(i = 0; i < size; i++) {
-               smpi_global->pending_send_request_queues[i]         = xbt_fifo_new();
-               smpi_global->pending_send_request_queues_mutexes[i] = SIMIX_mutex_init();
-               smpi_global->pending_recv_request_queues[i]         = xbt_fifo_new();
-               smpi_global->pending_recv_request_queues_mutexes[i] = SIMIX_mutex_init();
-               smpi_global->received_message_queues[i]             = xbt_fifo_new();
-               smpi_global->received_message_queues_mutexes[i]     = SIMIX_mutex_init();
-               smpi_global->timers[i]                              = xbt_os_timer_new();
-               smpi_global->timers_mutexes[i]                      = SIMIX_mutex_init();
-       }
+void smpi_process_simulated_start(void)
+{
+  smpi_process_data_t data = smpi_process_data();
 
+  data->simulated = SIMIX_get_clock();
 }
 
-void smpi_global_destroy()
+double smpi_process_simulated_elapsed(void)
 {
-       int i;
-
-       int size = SIMIX_host_get_number();
+  smpi_process_data_t data = smpi_process_data();
 
-       // start/stop
-       SIMIX_mutex_destroy(smpi_global->start_stop_mutex);
-       SIMIX_cond_destroy(smpi_global->start_stop_cond);
+  return SIMIX_get_clock() - data->simulated;
+}
 
-       // processes
-       xbt_free(smpi_global->sender_processes);
-       xbt_free(smpi_global->receiver_processes);
+MPI_Comm smpi_process_comm_self(void)
+{
+  smpi_process_data_t data = smpi_process_data();
 
-       // running hosts
-       SIMIX_mutex_destroy(smpi_global->running_hosts_count_mutex);
+  return data->comm_self;
+}
 
-       // mallocators
-       xbt_mallocator_free(smpi_global->request_mallocator);
-       xbt_mallocator_free(smpi_global->message_mallocator);
+void print_request(const char *message, MPI_Request request)
+{
+  XBT_DEBUG("%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);
+}
 
-       for(i = 0; i < size; i++) {
-               xbt_fifo_free(smpi_global->pending_send_request_queues[i]);
-               SIMIX_mutex_destroy(smpi_global->pending_send_request_queues_mutexes[i]);
-               xbt_fifo_free(smpi_global->pending_recv_request_queues[i]);
-               SIMIX_mutex_destroy(smpi_global->pending_recv_request_queues_mutexes[i]);
-               xbt_fifo_free(smpi_global->received_message_queues[i]);
-               SIMIX_mutex_destroy(smpi_global->received_message_queues_mutexes[i]);
-               xbt_os_timer_free(smpi_global->timers[i]);
-               SIMIX_mutex_destroy(smpi_global->timers_mutexes[i]);
-       }
+void smpi_global_init(void)
+{
+  int i;
+  MPI_Group group;
+  char name[MAILBOX_NAME_MAXLEN];
+
+  SIMIX_comm_set_copy_data_callback(&smpi_comm_copy_data_callback);
+  process_count = SIMIX_process_count();
+  process_data = xbt_new(smpi_process_data_t, process_count);
+  for (i = 0; i < process_count; i++) {
+    process_data[i] = xbt_new(s_smpi_process_data_t, 1);
+    process_data[i]->index = i;
+    process_data[i]->argc = NULL;
+    process_data[i]->argv = NULL;
+    process_data[i]->mailbox = SIMIX_req_rdv_create(get_mailbox_name(name, i));
+    process_data[i]->timer = xbt_os_timer_new();
+    group = smpi_group_new(1);
+    process_data[i]->comm_self = smpi_comm_new(group);
+    smpi_group_set_mapping(group, i, 0);
+  }
+  group = smpi_group_new(process_count);
+  MPI_COMM_WORLD = smpi_comm_new(group);
+  for (i = 0; i < process_count; i++) {
+    smpi_group_set_mapping(group, i, i);
+  }
+}
 
-       xbt_free(smpi_global->pending_send_request_queues);
-       xbt_free(smpi_global->pending_send_request_queues_mutexes);
-       xbt_free(smpi_global->pending_recv_request_queues);
-       xbt_free(smpi_global->pending_recv_request_queues_mutexes);
-       xbt_free(smpi_global->received_message_queues);
-       xbt_free(smpi_global->received_message_queues_mutexes);
-       xbt_free(smpi_global->timers);
-       xbt_free(smpi_global->timers_mutexes);
+void smpi_global_destroy(void)
+{
+  int count = smpi_process_count();
+  int i;
+
+  smpi_bench_destroy();
+  smpi_comm_destroy(MPI_COMM_WORLD);
+  MPI_COMM_WORLD = MPI_COMM_NULL;
+  for (i = 0; i < count; i++) {
+    smpi_comm_destroy(process_data[i]->comm_self);
+    xbt_os_timer_free(process_data[i]->timer);
+    SIMIX_req_rdv_destroy(process_data[i]->mailbox);
+    xbt_free(process_data[i]);
+  }
+  xbt_free(process_data);
+  process_data = NULL;
+}
 
-       xbt_free(smpi_global);
+/* Fortran specific stuff */
+/* With smpicc, the following weak symbols are used */
+/* With smpiff, the following weak symbols are replaced by those in libf2c */
+int __attribute__((weak)) xargc;
+char** __attribute__((weak)) xargv;
 
-       smpi_global = NULL;
+int __attribute__((weak)) main(int argc, char** argv) {
+   xargc = argc;
+   xargv = argv;
+   return MAIN__();
 }
 
-int smpi_run_simulation(int argc, char **argv)
+int MAIN__(void)
 {
-       xbt_fifo_item_t cond_item   = NULL;
-       smx_cond_t   cond           = NULL;
-       xbt_fifo_item_t action_item = NULL;
-       smx_action_t action         = NULL;
+  srand(SMPI_RAND_SEED);
+
+  double default_reference_speed = 20000.0;
+  xbt_cfg_register(&_surf_cfg_set, "smpi/running_power",
+                   "Power of the host running the simulation (in flop/s). Used to bench the operations.",
+                   xbt_cfgelm_double, &default_reference_speed, 1, 1, NULL,
+                   NULL);
 
-       xbt_fifo_t   actions_failed = xbt_fifo_new();
-       xbt_fifo_t   actions_done   = xbt_fifo_new();
+  int default_display_timing = 0;
+  xbt_cfg_register(&_surf_cfg_set, "smpi/display_timing",
+                   "Boolean indicating whether we should display the timing after simulation.",
+                   xbt_cfgelm_int, &default_display_timing, 1, 1, NULL,
+                   NULL);
 
-       srand(SMPI_RAND_SEED);
+  double default_threshold = 1e-6;
+  xbt_cfg_register(&_surf_cfg_set, "smpi/cpu_threshold",
+                   "Minimal computation time (in seconds) not discarded.",
+                   xbt_cfgelm_double, &default_threshold, 1, 1, NULL,
+                   NULL);
 
-       SIMIX_global_init(&argc, argv);
+  if(getenv("SMPI_PRETEND_CC") != NULL) {
+       /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the configuration tools */
+    return 0;
+  }
 
-       SIMIX_function_register("smpi_simulated_main", smpi_simulated_main);
-       SIMIX_function_register("smpi_sender",         smpi_sender);
-       SIMIX_function_register("smpi_receiver",       smpi_receiver);
+#ifdef HAVE_TRACING
+  TRACE_global_init(&xargc, xargv);
+#endif
 
-       // FIXME: ought to verify these files...
-       SIMIX_create_environment(argv[1]);
+  SIMIX_global_init(&xargc, xargv);
 
-       // must initialize globals between creating environment and launching app....
-       smpi_global_init();
+#ifdef HAVE_TRACING
+  TRACE_smpi_start();
+#endif
 
-       SIMIX_launch_application(argv[2]);
+  // parse the platform file: get the host list
+  SIMIX_create_environment(xargv[1]);
 
-       /* Prepare to display some more info when dying on Ctrl-C pressing */
-       // FIXME: doesn't work
-       //signal(SIGINT, inthandler);
+  SIMIX_function_register_default(smpi_simulated_main);
+  SIMIX_launch_application(xargv[2]);
 
-       /* Clean IO before the run */
-       fflush(stdout);
-       fflush(stderr);
+  smpi_global_init();
 
-       while (SIMIX_solve(actions_done, actions_failed) != -1.0) {
-               xbt_fifo_foreach(actions_failed, action_item, action, smx_action_t) {
-                       DEBUG1("** %s failed **", action->name);
-                       xbt_fifo_foreach(action->cond_list, cond_item, cond, smx_cond_t) {
-                               SIMIX_cond_broadcast(cond);
-                       }
-               }
-               xbt_fifo_foreach(actions_done, action_item, action, smx_action_t) {
-                       DEBUG1("** %s done **",action->name);
-                       xbt_fifo_foreach(action->cond_list, cond_item, cond, smx_cond_t) {
-                               SIMIX_cond_broadcast(cond);
-                       }
-               }
-       }
+  /* Clean IO before the run */
+  fflush(stdout);
+  fflush(stderr);
 
-       // FIXME: cleanup incomplete
-       xbt_fifo_free(actions_failed);
-       xbt_fifo_free(actions_done);
+  if (MC_IS_ENABLED)
+    MC_modelcheck();
+  else
+    SIMIX_run();
 
-       INFO1("simulation time %g", SIMIX_get_clock());
+  if (xbt_cfg_get_int(_surf_cfg_set, "smpi/display_timing"))
+    XBT_INFO("Simulation time: %g seconds.", SIMIX_get_clock());
 
-       smpi_global_destroy();
+  smpi_global_destroy();
 
-       SIMIX_clean();
+#ifdef HAVE_TRACING
+  TRACE_smpi_release();
+#endif
 
-       return 0;
+  SIMIX_clean();
+  return 0;
 }