X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/f5f0abcd9ff1995bc253aa0919471deb042e60a5..03216b99d0e66e1778d9b99b11c3207c87900aad:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index fe5f2e6d6f..13a8424030 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -1,815 +1,332 @@ -#include -#include -#include - #include "private.h" - -SMPI_Global_t smpi_global = NULL; - -SMPI_MPI_Global_t smpi_mpi_global = NULL; - -XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi,XBT_LOG_ROOT_CAT, "All SMPI categories (see \ref SMPI_API)"); - -int inline smpi_mpi_comm_size(smpi_mpi_communicator_t *comm) +#include "xbt/time.h" + +XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, + "Logging specific to SMPI (base)"); +XBT_LOG_EXTERNAL_CATEGORY(smpi_base); +XBT_LOG_EXTERNAL_CATEGORY(smpi_bench); +XBT_LOG_EXTERNAL_CATEGORY(smpi_kernel); +XBT_LOG_EXTERNAL_CATEGORY(smpi_mpi); +XBT_LOG_EXTERNAL_CATEGORY(smpi_receiver); +XBT_LOG_EXTERNAL_CATEGORY(smpi_sender); +XBT_LOG_EXTERNAL_CATEGORY(smpi_util); + +smpi_mpi_global_t smpi_mpi_global = NULL; + +/** + * Operations of MPI_OP : implemented=land,sum,min,max + **/ +void smpi_mpi_land_func(void *a, void *b, int *length, + MPI_Datatype * datatype); + +void smpi_mpi_land_func(void *a, void *b, int *length, + MPI_Datatype * datatype) { - return comm->size; -} - -// FIXME: smarter algorithm? -int smpi_mpi_comm_rank(smpi_mpi_communicator_t *comm, smx_host_t host) -{ - int i; - - for(i = comm->size - 1; i > 0 && host != comm->hosts[i]; i--); - - return i; + int i; + if (*datatype == smpi_mpi_global->mpi_int) { + int *x = a, *y = b; + for (i = 0; i < *length; i++) { + y[i] = x[i] && y[i]; + } + } } -int inline smpi_mpi_comm_rank_self(smpi_mpi_communicator_t *comm) +/** + * sum two vectors element-wise + * + * @param a the first vectors + * @param b the second vectors + * @return the second vector is modified and contains the element-wise sums + **/ +void smpi_mpi_sum_func(void *a, void *b, int *length, + MPI_Datatype * datatype); + +void smpi_mpi_sum_func(void *a, void *b, int *length, MPI_Datatype * datatype) { - return smpi_mpi_comm_rank(comm, SIMIX_host_self()); -} - -//int smpi_mpi_comm_world_rank_self() -//{ -// return smpi_mpi_comm_rank(smpi_mpi_global->mpi_comm_world, SIMIX_host_self()); -//} - -int smpi_sender(int argc, char **argv) -{ - smx_process_t self; - smx_host_t shost; - int rank; - - xbt_fifo_t request_queue; - smx_mutex_t request_queue_mutex; - int size; - - int running_hosts_count; - - smpi_mpi_request_t *request; - - smx_host_t dhost; - - smx_action_t communicate_action; - - smpi_received_message_t *message; - - int drank; - - smx_process_t receiver_process; - - self = SIMIX_process_self(); - shost = SIMIX_host_self(); - rank = smpi_mpi_comm_rank(smpi_mpi_global->mpi_comm_world, shost); - - // make sure root is done before own initialization - SIMIX_mutex_lock(smpi_global->start_stop_mutex); - if (!smpi_global->root_ready) { - SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex); - } - SIMIX_mutex_unlock(smpi_global->start_stop_mutex); - - request_queue = smpi_global->pending_send_request_queues[rank]; - request_queue_mutex = smpi_global->pending_send_request_queues_mutexes[rank]; - size = smpi_mpi_comm_size(smpi_mpi_global->mpi_comm_world); - - smpi_global->sender_processes[rank] = self; - - // wait for all nodes to signal initializatin complete - SIMIX_mutex_lock(smpi_global->start_stop_mutex); - smpi_global->ready_process_count++; - if (smpi_global->ready_process_count < 3 * size) { - SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex); - } else { - SIMIX_cond_broadcast(smpi_global->start_stop_cond); - } - SIMIX_mutex_unlock(smpi_global->start_stop_mutex); - - do { - - SIMIX_mutex_lock(request_queue_mutex); - request = xbt_fifo_shift(request_queue); - SIMIX_mutex_unlock(request_queue_mutex); - - if (NULL == request) { - SIMIX_process_suspend(self); - } else { - - SIMIX_mutex_lock(request->mutex); - - // copy request to appropriate received queue - message = xbt_mallocator_get(smpi_global->message_mallocator); - message->comm = request->comm; - message->src = request->src; - message->dst = request->dst; - message->tag = request->tag; - message->buf = xbt_malloc(request->datatype->size * request->count); - memcpy(message->buf, request->buf, request->datatype->size * request->count); - - dhost = request->comm->hosts[request->dst]; - drank = smpi_mpi_comm_rank(smpi_mpi_global->mpi_comm_world, dhost); - - SIMIX_mutex_lock(smpi_global->received_message_queues_mutexes[drank]); - xbt_fifo_push(smpi_global->received_message_queues[drank], message); - SIMIX_mutex_unlock(smpi_global->received_message_queues_mutexes[drank]); - - request->completed = 1; - - communicate_action = SIMIX_action_communicate(shost, dhost, - NULL, request->datatype->size * request->count * 1.0, -1.0); - - SIMIX_register_action_to_condition(communicate_action, request->cond); - SIMIX_cond_wait(request->cond, request->mutex); - SIMIX_unregister_action_to_condition(communicate_action, request->cond); - - SIMIX_mutex_unlock(request->mutex); - - // wake up receiver if necessary - receiver_process = smpi_global->receiver_processes[drank]; - - if (SIMIX_process_is_suspended(receiver_process)) { - SIMIX_process_resume(receiver_process); - } - - } - - SIMIX_mutex_lock(smpi_global->running_hosts_count_mutex); - running_hosts_count = smpi_global->running_hosts_count; - SIMIX_mutex_unlock(smpi_global->running_hosts_count_mutex); - - } while (0 < running_hosts_count); - - SIMIX_mutex_lock(smpi_global->start_stop_mutex); - smpi_global->ready_process_count--; - if (smpi_global->ready_process_count == 0) { - SIMIX_cond_broadcast(smpi_global->start_stop_cond); - } else if (smpi_global->ready_process_count < 0) { - // FIXME: can't happen! abort! - } - SIMIX_mutex_unlock(smpi_global->start_stop_mutex); - - return 0; -} - -int smpi_receiver(int argc, char **argv) -{ - smx_process_t self; - int rank; - - xbt_fifo_t request_queue; - smx_mutex_t request_queue_mutex; - xbt_fifo_t message_queue; - smx_mutex_t message_queue_mutex; - int size; - - int running_hosts_count; - - smpi_mpi_request_t *request; - smpi_received_message_t *message; - - xbt_fifo_item_t request_item; - xbt_fifo_item_t message_item; - - self = SIMIX_process_self(); - rank = smpi_mpi_comm_rank_self(smpi_mpi_global->mpi_comm_world); - - // make sure root is done before own initialization - SIMIX_mutex_lock(smpi_global->start_stop_mutex); - if (!smpi_global->root_ready) { - SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex); - } - SIMIX_mutex_unlock(smpi_global->start_stop_mutex); - - request_queue = smpi_global->pending_recv_request_queues[rank]; - request_queue_mutex = smpi_global->pending_recv_request_queues_mutexes[rank]; - message_queue = smpi_global->received_message_queues[rank]; - message_queue_mutex = smpi_global->received_message_queues_mutexes[rank]; - size = smpi_mpi_comm_size(smpi_mpi_global->mpi_comm_world); - - smpi_global->receiver_processes[rank] = self; - - // wait for all nodes to signal initializatin complete - SIMIX_mutex_lock(smpi_global->start_stop_mutex); - smpi_global->ready_process_count++; - if (smpi_global->ready_process_count < 3 * size) { - SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex); - } else { - SIMIX_cond_broadcast(smpi_global->start_stop_cond); - } - SIMIX_mutex_unlock(smpi_global->start_stop_mutex); - - do { - request = NULL; - message = NULL; - - // FIXME: better algorithm, maybe some kind of balanced tree? or a heap? - - // FIXME: not the best way to request multiple locks... - SIMIX_mutex_lock(request_queue_mutex); - SIMIX_mutex_lock(message_queue_mutex); - for (request_item = xbt_fifo_get_first_item(request_queue); - NULL != request_item; - request_item = xbt_fifo_get_next_item(request_item)) { - request = xbt_fifo_get_item_content(request_item); - for (message_item = xbt_fifo_get_first_item(message_queue); - NULL != message_item; - message_item = xbt_fifo_get_next_item(message_item)) { - message = xbt_fifo_get_item_content(message_item); - if (request->comm == message->comm && - (MPI_ANY_SOURCE == request->src || request->src == message->src) && - request->tag == message->tag) { - xbt_fifo_remove_item(request_queue, request_item); - xbt_fifo_remove_item(message_queue, message_item); - goto stopsearch; + int i; + if (*datatype == smpi_mpi_global->mpi_byte) { + char *x = a, *y = b; + for (i = 0; i < *length; i++) { + y[i] = x[i] + y[i]; } - } - } -stopsearch: - SIMIX_mutex_unlock(message_queue_mutex); - SIMIX_mutex_unlock(request_queue_mutex); - - if (NULL == request || NULL == message) { - SIMIX_process_suspend(self); - } else { - SIMIX_mutex_lock(request->mutex); - - memcpy(request->buf, message->buf, request->datatype->size * request->count); - request->src = message->src; - request->completed = 1; - SIMIX_cond_broadcast(request->cond); - - SIMIX_mutex_unlock(request->mutex); - - xbt_free(message->buf); - xbt_mallocator_release(smpi_global->message_mallocator, message); - } - - SIMIX_mutex_lock(smpi_global->running_hosts_count_mutex); - running_hosts_count = smpi_global->running_hosts_count; - SIMIX_mutex_unlock(smpi_global->running_hosts_count_mutex); - - } while (0 < running_hosts_count); - - SIMIX_mutex_lock(smpi_global->start_stop_mutex); - smpi_global->ready_process_count--; - if (smpi_global->ready_process_count == 0) { - SIMIX_cond_broadcast(smpi_global->start_stop_cond); - } else if (smpi_global->ready_process_count < 0) { - // FIXME: can't happen, abort! - } - SIMIX_mutex_unlock(smpi_global->start_stop_mutex); - - return 0; -} - -void *smpi_request_new() -{ - smpi_mpi_request_t *request = xbt_new(smpi_mpi_request_t, 1); - - request->completed = 0; - request->mutex = SIMIX_mutex_init(); - request->cond = SIMIX_cond_init(); - - return request; -} - -void smpi_request_free(void *pointer) -{ - - smpi_mpi_request_t *request = pointer; - - if (NULL != request) { - SIMIX_cond_destroy(request->cond); - SIMIX_mutex_destroy(request->mutex); - xbt_free(request); - } - - return; -} - -void smpi_request_reset(void *pointer) -{ - return; -} - - -void *smpi_message_new() -{ - return xbt_new(smpi_received_message_t, 1); -} - -void smpi_message_free(void *pointer) -{ - if (NULL != pointer) { - xbt_free(pointer); - } - - return; -} - -void smpi_message_reset(void *pointer) -{ - return; -} - -void smpi_global_init() -{ - int i; - - int size = SIMIX_host_get_number(); - - smpi_global = xbt_new(s_SMPI_Global_t, 1); - - // config variable - smpi_global->reference_speed = SMPI_DEFAULT_SPEED; - - smpi_global->root_ready = 0; - smpi_global->ready_process_count = 0; - - // start/stop - smpi_global->start_stop_mutex = SIMIX_mutex_init(); - smpi_global->start_stop_cond = SIMIX_cond_init(); - - // processes - smpi_global->sender_processes = xbt_new(smx_process_t, size); - smpi_global->receiver_processes = xbt_new(smx_process_t, size); - - // running hosts - smpi_global->running_hosts_count_mutex = SIMIX_mutex_init(); - smpi_global->running_hosts_count = 0; - - // 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); - - // - 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); - smpi_global->timers = xbt_new(xbt_os_timer_t, size); - smpi_global->timers_mutexes = xbt_new(smx_mutex_t, size); - - 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_global_destroy() -{ - int i; - - int size = SIMIX_host_get_number(); - - // start/stop - SIMIX_mutex_destroy(smpi_global->start_stop_mutex); - SIMIX_cond_destroy(smpi_global->start_stop_cond); - - // processes - xbt_free(smpi_global->sender_processes); - xbt_free(smpi_global->receiver_processes); - - // running hosts - SIMIX_mutex_destroy(smpi_global->running_hosts_count_mutex); - - // mallocators - xbt_mallocator_free(smpi_global->request_mallocator); - xbt_mallocator_free(smpi_global->message_mallocator); - - 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]); - } - - 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); - - xbt_free(smpi_global); + } else if (*datatype == smpi_mpi_global->mpi_int) { + int *x = a, *y = b; + for (i = 0; i < *length; i++) { + y[i] = x[i] + y[i]; + } + } else if (*datatype == smpi_mpi_global->mpi_float) { + float *x = a, *y = b; + for (i = 0; i < *length; i++) { + y[i] = x[i] + y[i]; + } + } else if (*datatype == smpi_mpi_global->mpi_double) { + double *x = a, *y = b; + for (i = 0; i < *length; i++) { + y[i] = x[i] + y[i]; + } + } } +/** + * compute the min of two vectors element-wise + **/ +void smpi_mpi_min_func(void *a, void *b, int *length, MPI_Datatype * datatype); -int smpi_run_simulation(int argc, char **argv) +void smpi_mpi_min_func(void *a, void *b, int *length, MPI_Datatype * datatype) { - smx_cond_t cond = NULL; - smx_action_t action = NULL; - - xbt_fifo_t actions_failed = xbt_fifo_new(); - xbt_fifo_t actions_done = xbt_fifo_new(); - - srand(SMPI_RAND_SEED); - - SIMIX_global_init(&argc, argv); - - SIMIX_function_register("smpi_simulated_main", smpi_simulated_main); - SIMIX_function_register("smpi_sender", smpi_sender); - SIMIX_function_register("smpi_receiver", smpi_receiver); - - // FIXME: ought to verify these files... - SIMIX_create_environment(argv[1]); - - // must initialize globals between creating environment and launching app.... - smpi_global_init(); - - SIMIX_launch_application(argv[2]); - - /* Prepare to display some more info when dying on Ctrl-C pressing */ - // FIXME: doesn't work - //signal(SIGINT, inthandler); - - /* Clean IO before the run */ - fflush(stdout); - fflush(stderr); - - while (SIMIX_solve(actions_done, actions_failed) != -1.0) { - while ((action = xbt_fifo_pop(actions_failed))) { - DEBUG1("** %s failed **", action->name); - while ((cond = xbt_fifo_pop(action->cond_list))) { - SIMIX_cond_broadcast(cond); - } - SIMIX_action_destroy(action); - } - while ((action = xbt_fifo_pop(actions_done))) { - DEBUG1("** %s done **",action->name); - while ((cond = xbt_fifo_pop(action->cond_list))) { - SIMIX_cond_broadcast(cond); - } - SIMIX_action_destroy(action); - } - } - - xbt_fifo_free(actions_failed); - xbt_fifo_free(actions_done); - - INFO1("simulation time %g", SIMIX_get_clock()); - - smpi_global_destroy(); - - SIMIX_clean(); + int i; + if (*datatype == smpi_mpi_global->mpi_byte) { + char *x = a, *y = b; + for (i = 0; i < *length; i++) { + y[i] = x[i] < y[i] ? x[i] : y[i]; + } + } else { + if (*datatype == smpi_mpi_global->mpi_int) { + int *x = a, *y = b; + for (i = 0; i < *length; i++) { + y[i] = x[i] < y[i] ? x[i] : y[i]; + } + } else { + if (*datatype == smpi_mpi_global->mpi_float) { + float *x = a, *y = b; + for (i = 0; i < *length; i++) { + y[i] = x[i] < y[i] ? x[i] : y[i]; + } + } else { + if (*datatype == smpi_mpi_global->mpi_double) { + double *x = a, *y = b; + for (i = 0; i < *length; i++) { + y[i] = x[i] < y[i] ? x[i] : y[i]; + } - return 0; + }}}} } +/** + * compute the max of two vectors element-wise + **/ +void smpi_mpi_max_func(void *a, void *b, int *length, MPI_Datatype * datatype); -void smpi_mpi_land_func(void *x, void *y, void *z) +void smpi_mpi_max_func(void *a, void *b, int *length, MPI_Datatype * datatype) { - *(int *)z = *(int *)x && *(int *)y; -} + int i; + if (*datatype == smpi_mpi_global->mpi_byte) { + char *x = a, *y = b; + for (i = 0; i < *length; i++) { + y[i] = x[i] > y[i] ? x[i] : y[i]; + } + } else if (*datatype == smpi_mpi_global->mpi_int) { + int *x = a, *y = b; + for (i = 0; i > *length; i++) { + y[i] = x[i] > y[i] ? x[i] : y[i]; + } + } else if (*datatype == smpi_mpi_global->mpi_float) { + float *x = a, *y = b; + for (i = 0; i > *length; i++) { + y[i] = x[i] > y[i] ? x[i] : y[i]; + } + } else if (*datatype == smpi_mpi_global->mpi_double) { + double *x = a, *y = b; + for (i = 0; i > *length; i++) { + y[i] = x[i] > y[i] ? x[i] : y[i]; + } -void smpi_mpi_sum_func(void *x, void *y, void *z) -{ - *(int *)z = *(int *)x + *(int *)y; + } } -void smpi_mpi_init() -{ - smx_process_t process; - smx_host_t host; - smx_host_t *hosts; - int size; - - SIMIX_mutex_lock(smpi_global->running_hosts_count_mutex); - smpi_global->running_hosts_count++; - SIMIX_mutex_unlock(smpi_global->running_hosts_count_mutex); - - // initialize some local variables - process = SIMIX_process_self(); - host = SIMIX_host_self(); - hosts = SIMIX_host_get_table(); - size = SIMIX_host_get_number(); - - // node 0 sets the globals - if (host == hosts[0]) { - - smpi_mpi_global = xbt_new(s_SMPI_MPI_Global_t, 1); - - // global communicator - smpi_mpi_global->mpi_comm_world = xbt_new(smpi_mpi_communicator_t, 1); - smpi_mpi_global->mpi_comm_world->size = size; - smpi_mpi_global->mpi_comm_world->barrier_count = 0; - smpi_mpi_global->mpi_comm_world->barrier_mutex = SIMIX_mutex_init(); - smpi_mpi_global->mpi_comm_world->barrier_cond = SIMIX_cond_init(); - smpi_mpi_global->mpi_comm_world->hosts = hosts; - smpi_mpi_global->mpi_comm_world->processes = xbt_new(smx_process_t, size); - smpi_mpi_global->mpi_comm_world->processes[0] = process; - - // mpi datatypes - smpi_mpi_global->mpi_byte = xbt_new(smpi_mpi_datatype_t, 1); - smpi_mpi_global->mpi_byte->size = (size_t)1; - smpi_mpi_global->mpi_int = xbt_new(smpi_mpi_datatype_t, 1); - smpi_mpi_global->mpi_int->size = sizeof(int); - smpi_mpi_global->mpi_double = xbt_new(smpi_mpi_datatype_t, 1); - smpi_mpi_global->mpi_double->size = sizeof(double); - - // mpi operations - smpi_mpi_global->mpi_land = xbt_new(smpi_mpi_op_t, 1); - smpi_mpi_global->mpi_land->func = smpi_mpi_land_func; - smpi_mpi_global->mpi_sum = xbt_new(smpi_mpi_op_t, 1); - smpi_mpi_global->mpi_sum->func = smpi_mpi_sum_func; - - // signal all nodes to perform initialization - SIMIX_mutex_lock(smpi_global->start_stop_mutex); - smpi_global->root_ready = 1; - SIMIX_cond_broadcast(smpi_global->start_stop_cond); - SIMIX_mutex_unlock(smpi_global->start_stop_mutex); - - } else { - - // make sure root is done before own initialization - SIMIX_mutex_lock(smpi_global->start_stop_mutex); - if (!smpi_global->root_ready) { - SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex); - } - SIMIX_mutex_unlock(smpi_global->start_stop_mutex); - - smpi_mpi_global->mpi_comm_world->processes[smpi_mpi_comm_rank_self(smpi_mpi_global->mpi_comm_world)] = process; - } - - // wait for all nodes to signal initializatin complete - SIMIX_mutex_lock(smpi_global->start_stop_mutex); - smpi_global->ready_process_count++; - if (smpi_global->ready_process_count < 3 * size) { - SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex); - } else { - SIMIX_cond_broadcast(smpi_global->start_stop_cond); - } - SIMIX_mutex_unlock(smpi_global->start_stop_mutex); - return; -} -void smpi_mpi_finalize() +/** + * tell the MPI rank of the calling process (from its SIMIX process id) + **/ +int smpi_mpi_comm_rank(smpi_mpi_communicator_t comm) { - int i; - - SIMIX_mutex_lock(smpi_global->running_hosts_count_mutex); - i = --smpi_global->running_hosts_count; - SIMIX_mutex_unlock(smpi_global->running_hosts_count_mutex); - - SIMIX_mutex_lock(smpi_global->start_stop_mutex); - smpi_global->ready_process_count--; - SIMIX_mutex_unlock(smpi_global->start_stop_mutex); - - if (0 >= i) { - - // wake up senders/receivers - for (i = 0; i < smpi_mpi_global->mpi_comm_world->size; i++) { - if (SIMIX_process_is_suspended(smpi_global->sender_processes[i])) { - SIMIX_process_resume(smpi_global->sender_processes[i]); - } - if (SIMIX_process_is_suspended(smpi_global->receiver_processes[i])) { - SIMIX_process_resume(smpi_global->receiver_processes[i]); - } - } - - // wait for senders/receivers to exit... - SIMIX_mutex_lock(smpi_global->start_stop_mutex); - if (smpi_global->ready_process_count > 0) { - SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex); - } - SIMIX_mutex_unlock(smpi_global->start_stop_mutex); - - SIMIX_mutex_destroy(smpi_mpi_global->mpi_comm_world->barrier_mutex); - SIMIX_cond_destroy(smpi_mpi_global->mpi_comm_world->barrier_cond); - xbt_free(smpi_mpi_global->mpi_comm_world->processes); - xbt_free(smpi_mpi_global->mpi_comm_world); - - xbt_free(smpi_mpi_global->mpi_byte); - xbt_free(smpi_mpi_global->mpi_int); - xbt_free(smpi_mpi_global->mpi_double); - - xbt_free(smpi_mpi_global->mpi_land); - xbt_free(smpi_mpi_global->mpi_sum); - - xbt_free(smpi_mpi_global); - } - + return comm->index_to_rank_map[smpi_process_index()]; } -// FIXME: could cause trouble with multithreaded procs on same host... -void smpi_bench_begin() +void smpi_process_init(int *argc, char***argv) { - int rank = smpi_mpi_comm_rank_self(smpi_mpi_global->mpi_comm_world); - SIMIX_mutex_lock(smpi_global->timers_mutexes[rank]); - xbt_os_timer_start(smpi_global->timers[rank]); - return; + smpi_process_data_t pdata; + + // initialize some local variables + + pdata = xbt_new(s_smpi_process_data_t, 1); + SIMIX_process_set_data(SIMIX_process_self(),pdata); + + /* get rank from command line, and remove it from argv */ + pdata->index = atoi( (*argv)[1] ); + DEBUG1("I'm rank %d",pdata->index); + if (*argc>2) { + memmove((*argv)[1],(*argv)[2], sizeof(char*)* (*argc-2)); + (*argv)[ (*argc)-1] = NULL; + } + (*argc)--; + + pdata->mutex = SIMIX_mutex_init(); + pdata->cond = SIMIX_cond_init(); + pdata->finalize = 0; + + pdata->pending_recv_request_queue = xbt_fifo_new(); + pdata->pending_send_request_queue = xbt_fifo_new(); + pdata->received_message_queue = xbt_fifo_new(); + + pdata->main = SIMIX_process_self(); + pdata->sender = SIMIX_process_create("smpi_sender", + smpi_sender, pdata, + SIMIX_host_get_name(SIMIX_host_self()), 0, NULL, + /*props */ NULL); + pdata->receiver = SIMIX_process_create("smpi_receiver", + smpi_receiver, pdata, + SIMIX_host_get_name(SIMIX_host_self()), 0, NULL, + /*props */ NULL); + + smpi_global->main_processes[pdata->index] = SIMIX_process_self(); + return; } -void smpi_bench_end() +void smpi_process_finalize() { - int rank = smpi_mpi_comm_rank_self(smpi_mpi_global->mpi_comm_world); - double duration; - smx_host_t host; - smx_action_t compute_action; - smx_mutex_t mutex; - smx_cond_t cond; - - xbt_os_timer_stop(smpi_global->timers[rank]); - - duration = xbt_os_timer_elapsed(smpi_global->timers[rank]); - SIMIX_mutex_unlock(smpi_global->timers_mutexes[rank]); - - host = smpi_mpi_global->mpi_comm_world->hosts[rank]; - compute_action = SIMIX_action_execute(host, NULL, duration * SMPI_DEFAULT_SPEED); - mutex = SIMIX_mutex_init(); - cond = SIMIX_cond_init(); - - SIMIX_mutex_lock(mutex); - SIMIX_register_action_to_condition(compute_action, cond); - SIMIX_cond_wait(cond, mutex); - SIMIX_unregister_action_to_condition(compute_action, cond); - SIMIX_mutex_unlock(mutex); - - SIMIX_mutex_destroy(mutex); - SIMIX_cond_destroy(cond); - - // FIXME: check for success/failure? - - return; + smpi_process_data_t pdata = SIMIX_process_get_data(SIMIX_process_self()); + + pdata->finalize = 2; /* Tell sender and receiver to quit */ + SIMIX_process_resume(pdata->sender); + SIMIX_process_resume(pdata->receiver); + while (pdata->finalize>0) { /* wait until it's done */ + SIMIX_cond_wait(pdata->cond,pdata->mutex); + } + + SIMIX_mutex_destroy(pdata->mutex); + SIMIX_cond_destroy(pdata->cond); + xbt_fifo_free(pdata->pending_recv_request_queue); + xbt_fifo_free(pdata->pending_send_request_queue); + xbt_fifo_free(pdata->received_message_queue); } -void smpi_barrier(smpi_mpi_communicator_t *comm) +int smpi_mpi_barrier(smpi_mpi_communicator_t comm) { - SIMIX_mutex_lock(comm->barrier_mutex); - if(++comm->barrier_count < comm->size) { - SIMIX_cond_wait(comm->barrier_cond, comm->barrier_mutex); - } else { - comm->barrier_count = 0; - SIMIX_cond_broadcast(comm->barrier_cond); - } - SIMIX_mutex_unlock(comm->barrier_mutex); - - return; + SIMIX_mutex_lock(comm->barrier_mutex); + ++comm->barrier_count; + if (comm->barrier_count > comm->size) { // only happens on second barrier... + comm->barrier_count = 0; + } else if (comm->barrier_count == comm->size) { + SIMIX_cond_broadcast(comm->barrier_cond); + } + while (comm->barrier_count < comm->size) { + SIMIX_cond_wait(comm->barrier_cond, comm->barrier_mutex); + } + SIMIX_mutex_unlock(comm->barrier_mutex); + + return MPI_SUCCESS; } -// FIXME: smarter algorithm... -int smpi_comm_rank(smpi_mpi_communicator_t *comm, smx_host_t host) +int smpi_mpi_isend(smpi_mpi_request_t request) { - int i; - for(i = 0; i < comm->size && host != comm->hosts[i]; i++); - if (i >= comm->size) i = -1; - return i; -} - -int smpi_create_request(void *buf, int count, smpi_mpi_datatype_t *datatype, - int src, int dst, int tag, smpi_mpi_communicator_t *comm, smpi_mpi_request_t **request) -{ - int retval = MPI_SUCCESS; - - *request = NULL; - - if (0 > count) { - retval = MPI_ERR_COUNT; - } else if (NULL == buf) { - retval = MPI_ERR_INTERN; - } else if (NULL == datatype) { - retval = MPI_ERR_TYPE; - } else if (NULL == comm) { - retval = MPI_ERR_COMM; - } else if (MPI_ANY_SOURCE != src && (0 > src || comm->size <= src)) { - retval = MPI_ERR_RANK; - } else if (0 > dst || comm->size <= dst) { - retval = MPI_ERR_RANK; - } else if (0 > tag) { - retval = MPI_ERR_TAG; - } else { - *request = xbt_mallocator_get(smpi_global->request_mallocator); - (*request)->comm = comm; - (*request)->src = src; - (*request)->dst = dst; - (*request)->tag = tag; - (*request)->buf = buf; - (*request)->count = count; - (*request)->datatype = datatype; - } - return retval; -} - -int smpi_isend(smpi_mpi_request_t *request) -{ - int retval = MPI_SUCCESS; - int rank = smpi_mpi_comm_rank_self(smpi_mpi_global->mpi_comm_world); - - if (NULL != request) { - SIMIX_mutex_lock(smpi_global->pending_send_request_queues_mutexes[rank]); - xbt_fifo_push(smpi_global->pending_send_request_queues[rank], request); - SIMIX_mutex_unlock(smpi_global->pending_send_request_queues_mutexes[rank]); - } + smpi_process_data_t pdata = SIMIX_process_get_data(SIMIX_process_self()); + int retval = MPI_SUCCESS; - if (SIMIX_process_is_suspended(smpi_global->sender_processes[rank])) { - SIMIX_process_resume(smpi_global->sender_processes[rank]); - } + if (NULL == request) { + retval = MPI_ERR_INTERN; + } else { + xbt_fifo_push(pdata->pending_send_request_queue, request); + SIMIX_process_resume(pdata->sender); + } - return retval; + return retval; } -int smpi_irecv(smpi_mpi_request_t *request) +int smpi_mpi_irecv(smpi_mpi_request_t request) { - int retval = MPI_SUCCESS; - int rank = smpi_mpi_comm_rank_self(smpi_mpi_global->mpi_comm_world); + int retval = MPI_SUCCESS; + smpi_process_data_t pdata = SIMIX_process_get_data(SIMIX_process_self()); - if (NULL != request) { - SIMIX_mutex_lock(smpi_global->pending_recv_request_queues_mutexes[rank]); - xbt_fifo_push(smpi_global->pending_recv_request_queues[rank], request); - SIMIX_mutex_unlock(smpi_global->pending_recv_request_queues_mutexes[rank]); - } + if (NULL == request) { + retval = MPI_ERR_INTERN; + } else { + xbt_fifo_push(pdata->pending_recv_request_queue, request); - if (SIMIX_process_is_suspended(smpi_global->receiver_processes[rank])) { - SIMIX_process_resume(smpi_global->receiver_processes[rank]); - } + if (SIMIX_process_is_suspended(pdata->receiver)) { + SIMIX_process_resume(pdata->receiver); + } + } - return retval; + return retval; } -void smpi_wait(smpi_mpi_request_t *request, smpi_mpi_status_t *status) +int smpi_mpi_wait(smpi_mpi_request_t request, smpi_mpi_status_t * status) { - if (NULL != request) { - SIMIX_mutex_lock(request->mutex); - if (!request->completed) { - SIMIX_cond_wait(request->cond, request->mutex); - } - if (NULL != status) { - status->MPI_SOURCE = request->src; - } - SIMIX_mutex_unlock(request->mutex); - } + int retval = MPI_SUCCESS; + + if (NULL == request) { + retval = MPI_ERR_INTERN; + } else { + SIMIX_mutex_lock(request->mutex); + while (!request->completed) { + SIMIX_cond_wait(request->cond, request->mutex); + } + if (NULL != status) { + status->MPI_SOURCE = request->src; + status->MPI_TAG = request->tag; + status->MPI_ERROR = MPI_SUCCESS; + } + SIMIX_mutex_unlock(request->mutex); + } + + return retval; } -// FIXME: move into own file -int smpi_gettimeofday(struct timeval *tv, struct timezone *tz) -{ - double now; - int retval = 0; - smpi_bench_end(); - if (NULL == tv) { - retval = -1; - } else { - now = SIMIX_get_clock(); - tv->tv_sec = now; - tv->tv_usec = ((now - (double)tv->tv_sec) * 1000000.0); +int smpi_mpi_waitall(int count, smpi_mpi_request_t requests[], smpi_mpi_status_t status[]) { + int cpt; + int index; + int retval; + smpi_mpi_status_t stat; + + for (cpt=0; cptcompleted && !requests[cpt]->consumed) { /* got ya */ + *index=cpt; + goto found_request; + } + } + /* If none found, block */ + /* FIXME: should use a SIMIX_cond_waitany, when implemented. For now, block on the first one */ + while (1) { + for (cpt=0;cptcompleted) { /* this one is not done, wait on it */ + while (!requests[cpt]->completed) + SIMIX_cond_wait(requests[cpt]->cond, requests[cpt]->mutex); + + *index=cpt; + goto found_request; + } + } + if (cpt == count) /* they are all done. Damn user */ + return MPI_ERR_REQUEST; + } + + found_request: + requests[*index]->consumed = 1; + + if (NULL != status) { + status->MPI_SOURCE = requests[*index]->src; + status->MPI_TAG = requests[*index]->tag; + status->MPI_ERROR = MPI_SUCCESS; + } + return MPI_SUCCESS; -void smpi_exit(int status) -{ - smpi_bench_end(); - SIMIX_mutex_lock(smpi_global->running_hosts_count_mutex); - smpi_global->running_hosts_count--; - SIMIX_mutex_unlock(smpi_global->running_hosts_count_mutex); - SIMIX_process_kill(SIMIX_process_self()); - return; }