-#include <stdio.h>
-#include <signal.h>
-#include <sys/time.h>
+/* 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 "private.h"
+#include "xbt/virtu.h"
+#include "mc/mc.h"
+#include "xbt/replay.h"
+#include <errno.h>
+#include "simix/smx_private.h"
+#include "surf/surf.h"
+#include "simgrid/sg_config.h"
+
+
+XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
+
+
+static int match_recv(void* a, void* b, smx_action_t ignored) {
+ MPI_Request ref = (MPI_Request)a;
+ MPI_Request req = (MPI_Request)b;
+ XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
+
+ xbt_assert(ref, "Cannot match recv against null reference");
+ xbt_assert(req, "Cannot match recv against null request");
+ if((ref->src == MPI_ANY_SOURCE || req->src == ref->src)
+ && (ref->tag == MPI_ANY_TAG || req->tag == ref->tag)){
+ //we match, we can transfer some values
+ // FIXME : move this to the copy function ?
+ if(ref->src == MPI_ANY_SOURCE)ref->real_src = req->src;
+ if(ref->tag == MPI_ANY_TAG)ref->real_tag = req->tag;
+ if(ref->real_size < req->real_size) ref->truncated = 1;
+ if(req->detached==1){
+ ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
+ }
+ return 1;
+ }else return 0;
+}
-SMPI_Global_t smpi_global = NULL;
+static int match_send(void* a, void* b,smx_action_t ignored) {
+ MPI_Request ref = (MPI_Request)a;
+ MPI_Request req = (MPI_Request)b;
+ XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
+ xbt_assert(ref, "Cannot match send against null reference");
+ xbt_assert(req, "Cannot match send against null request");
+
+ if((req->src == MPI_ANY_SOURCE || req->src == ref->src)
+ && (req->tag == MPI_ANY_TAG || req->tag == ref->tag))
+ {
+ if(req->src == MPI_ANY_SOURCE)req->real_src = ref->src;
+ if(req->tag == MPI_ANY_TAG)req->real_tag = ref->tag;
+ if(req->real_size < ref->real_size) req->truncated = 1;
+ if(ref->detached==1){
+ req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
+ }
+
+ return 1;
+ } else return 0;
+}
-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)");
+typedef struct s_smpi_factor *smpi_factor_t;
+typedef struct s_smpi_factor {
+ long factor;
+ int nb_values;
+ double values[4];//arbitrary set to 4
+} s_smpi_factor_t;
+xbt_dynar_t smpi_os_values = NULL;
+xbt_dynar_t smpi_or_values = NULL;
+xbt_dynar_t smpi_ois_values = NULL;
-int inline smpi_mpi_comm_size(smpi_mpi_communicator_t *comm)
-{
- return comm->size;
-}
+// Methods used to parse and store the values for timing injections in smpi
+// These are taken from surf/network.c and generalized to have more factors
+// These methods should be merged with those in surf/network.c (moved somewhere in xbt ?)
-// FIXME: smarter algorithm?
-int smpi_mpi_comm_rank(smpi_mpi_communicator_t *comm, smx_host_t host)
+static int factor_cmp(const void *pa, const void *pb)
{
- int i;
+ return (((s_smpi_factor_t*)pa)->factor > ((s_smpi_factor_t*)pb)->factor);
+}
- for(i = comm->size - 1; i > 0 && host != comm->hosts[i]; i--);
- return i;
+static xbt_dynar_t parse_factor(const char *smpi_coef_string)
+{
+ char *value = NULL;
+ unsigned int iter = 0;
+ s_smpi_factor_t fact;
+ int i=0;
+ xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL;
+
+ smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL);
+ radical_elements = xbt_str_split(smpi_coef_string, ";");
+ xbt_dynar_foreach(radical_elements, iter, value) {
+ fact.nb_values=0;
+ radical_elements2 = xbt_str_split(value, ":");
+ if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5)
+ xbt_die("Malformed radical for smpi factor!");
+ for(i =0; i<xbt_dynar_length(radical_elements2);i++ ){
+ if (i==0){
+ fact.factor = atol(xbt_dynar_get_as(radical_elements2, i, char *));
+ }else{
+ fact.values[fact.nb_values] = atof(xbt_dynar_get_as(radical_elements2, i, char *));
+ fact.nb_values++;
+ }
+ }
+
+ xbt_dynar_push_as(smpi_factor, s_smpi_factor_t, fact);
+ XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
+ xbt_dynar_free(&radical_elements2);
+ }
+ xbt_dynar_free(&radical_elements);
+ iter=0;
+ xbt_dynar_sort(smpi_factor, &factor_cmp);
+ xbt_dynar_foreach(smpi_factor, iter, fact) {
+ XBT_DEBUG("smpi_factor:\t%ld : %d values, first: %f", fact.factor, fact.nb_values ,fact.values[0]);
+ }
+ return smpi_factor;
}
-int inline smpi_mpi_comm_rank_self(smpi_mpi_communicator_t *comm)
+static double smpi_os(double size)
{
- return smpi_mpi_comm_rank(comm, SIMIX_host_self());
+ if (!smpi_os_values)
+ smpi_os_values =
+ parse_factor(sg_cfg_get_string("smpi/os"));
+
+ unsigned int iter = 0;
+ s_smpi_factor_t fact;
+ double current=0.0;
+ xbt_dynar_foreach(smpi_os_values, iter, fact) {
+ if (size <= fact.factor) {
+ XBT_DEBUG("os : %lf <= %ld return %f", size, fact.factor, current);
+ return current;
+ }else{
+ current=fact.values[0]+fact.values[1]*size;
+ }
+ }
+ XBT_DEBUG("os : %lf > %ld return %f", size, fact.factor, current);
+
+ return current;
}
-int smpi_sender(int argc, char **argv)
+static double smpi_ois(double size)
{
- 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 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 {
-
- message = xbt_mallocator_get(smpi_global->message_mallocator);
-
- SIMIX_mutex_lock(request->simdata->mutex);
-
- 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;
-
- action = SIMIX_action_communicate(shost, dhost, NULL, request->datatype->size * request->count * 1.0, -1.0);
-
- SIMIX_register_action_to_condition(action, request->simdata->cond);
-
- SIMIX_cond_wait(request->simdata->cond, request->simdata->mutex);
-
- SIMIX_mutex_unlock(request->simdata->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;
+ if (!smpi_ois_values)
+ smpi_ois_values =
+ parse_factor(sg_cfg_get_string("smpi/ois"));
+
+ unsigned int iter = 0;
+ s_smpi_factor_t fact;
+ double current=0.0;
+ xbt_dynar_foreach(smpi_ois_values, iter, fact) {
+ if (size <= fact.factor) {
+ XBT_DEBUG("ois : %lf <= %ld return %f", size, fact.factor, current);
+ return current;
+ }else{
+ current=fact.values[0]+fact.values[1]*size;
+ }
+ }
+ XBT_DEBUG("ois : %lf > %ld return %f", size, fact.factor, current);
+
+ return current;
}
-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;
- }
- }
- }
-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->simdata->mutex);
-
- memcpy(request->buf, message->buf, request->datatype->size * request->count);
- request->src = message->src;
- request->completed = 1;
- SIMIX_cond_broadcast(request->simdata->cond);
-
- SIMIX_mutex_unlock(request->simdata->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;
+static double smpi_or(double size)
+{
+ if (!smpi_or_values)
+ smpi_or_values =
+ parse_factor(sg_cfg_get_string("smpi/or"));
+
+ unsigned int iter = 0;
+ s_smpi_factor_t fact;
+ double current=0.0;
+ xbt_dynar_foreach(smpi_or_values, iter, fact) {
+ if (size <= fact.factor) {
+ XBT_DEBUG("or : %lf <= %ld return %f", size, fact.factor, current);
+ return current;
+ }else
+ current=fact.values[0]+fact.values[1]*size;
+ }
+ XBT_DEBUG("or : %lf > %ld return %f", size, fact.factor, current);
+
+ return current;
}
-void *smpi_request_new()
+static MPI_Request build_request(void *buf, int count,
+ MPI_Datatype datatype, int src, int dst,
+ int tag, MPI_Comm comm, unsigned flags)
{
- smpi_mpi_request_t *request = xbt_new(smpi_mpi_request_t, 1);
+ MPI_Request request;
+
+ void *old_buf = NULL;
+
+ request = xbt_new(s_smpi_mpi_request_t, 1);
+
+ s_smpi_subtype_t *subtype = datatype->substruct;
+
+ if(datatype->has_subtype == 1){
+ // This part handles the problem of non-contiguous memory
+ old_buf = buf;
+ buf = xbt_malloc(count*smpi_datatype_size(datatype));
+ if (flags & SEND) {
+ subtype->serialize(old_buf, buf, count, datatype->substruct);
+ }
+ }
+
+ request->buf = buf;
+ // This part handles the problem of non-contiguous memory (for the
+ // unserialisation at the reception)
+ request->old_buf = old_buf;
+ request->old_type = datatype;
+
+ request->size = smpi_datatype_size(datatype) * count;
+ request->src = src;
+ request->dst = dst;
+ request->tag = tag;
+ request->comm = comm;
+ request->action = NULL;
+ request->flags = flags;
+ request->detached = 0;
+ request->detached_sender = NULL;
+
+ request->truncated = 0;
+ request->real_size = 0;
+ request->real_tag = 0;
+
+ request->refcount=1;
+#ifdef HAVE_TRACING
+ request->send = 0;
+ request->recv = 0;
+#endif
+ if (flags & SEND) smpi_datatype_unuse(datatype);
+
+ return request;
+}
- request->completed = 0;
- request->simdata = xbt_new(s_smpi_mpi_request_simdata_t, 1);
- request->simdata->mutex = SIMIX_mutex_init();
- request->simdata->cond = SIMIX_cond_init();
- return request;
+void smpi_empty_status(MPI_Status * status)
+{
+ if(status != MPI_STATUS_IGNORE) {
+ status->MPI_SOURCE = MPI_ANY_SOURCE;
+ status->MPI_TAG = MPI_ANY_TAG;
+ status->MPI_ERROR = MPI_SUCCESS;
+ status->count=0;
+ }
}
-void smpi_request_free(void *pointer)
+void smpi_action_trace_run(char *path)
{
-
- smpi_mpi_request_t *request = pointer;
-
- if (NULL != request) {
- SIMIX_cond_destroy(request->simdata->cond);
- SIMIX_mutex_destroy(request->simdata->mutex);
- xbt_free(request->simdata);
- xbt_free(request);
- }
-
- return;
+ char *name;
+ xbt_dynar_t todo;
+ xbt_dict_cursor_t cursor;
+
+ action_fp=NULL;
+ if (path) {
+ action_fp = fopen(path, "r");
+ xbt_assert(action_fp != NULL, "Cannot open %s: %s", path,
+ strerror(errno));
+ }
+
+ if (!xbt_dict_is_empty(action_queues)) {
+ XBT_WARN
+ ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
+
+
+ xbt_dict_foreach(action_queues, cursor, name, todo) {
+ XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
+ }
+ }
+
+ if (path)
+ fclose(action_fp);
+ xbt_dict_free(&action_queues);
+ action_queues = xbt_dict_new_homogeneous(NULL);
}
-void smpi_request_reset(void *pointer)
+static void smpi_mpi_request_free_voidp(void* request)
{
- return;
+ MPI_Request req = request;
+ smpi_mpi_request_free(&req);
}
-
-void *smpi_message_new()
+/* MPI Low level calls */
+MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
+ int dst, int tag, MPI_Comm comm)
{
- return xbt_new(smpi_received_message_t, 1);
+ MPI_Request request =
+ build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
+ comm, PERSISTENT | SEND);
+ request->refcount++;
+ return request;
}
-void smpi_message_free(void *pointer)
+MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
+ int dst, int tag, MPI_Comm comm)
{
- if (NULL != pointer) {
- xbt_free(pointer);
- }
-
- return;
+ MPI_Request request =
+ build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
+ comm, PERSISTENT | SSEND | SEND);
+ request->refcount++;
+ return request;
}
-void smpi_message_reset(void *pointer)
+MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
+ int src, int tag, MPI_Comm comm)
{
- return;
+ MPI_Request request =
+ build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
+ comm, PERSISTENT | RECV);
+ request->refcount++;
+ return request;
}
-void smpi_global_init()
+void smpi_mpi_start(MPI_Request request)
{
- 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();
- }
+ smx_rdv_t mailbox;
+
+ xbt_assert(!request->action,
+ "Cannot (re)start a non-finished communication");
+ if(request->flags & RECV) {
+ print_request("New recv", request);
+ if (request->size < sg_cfg_get_int("smpi/async_small_thres"))
+ mailbox = smpi_process_mailbox_small();
+ else
+ mailbox = smpi_process_mailbox();
+ // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
+ request->real_size=request->size;
+ smpi_datatype_use(request->old_type);
+ request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request);
+
+ //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
+ double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
+ if(sleeptime!=0.0){
+ simcall_process_sleep(sleeptime);
+ XBT_DEBUG("receiving size of %zu : sleep %lf ", request->size, smpi_or(request->size));
+ }
+
+ } else {
+
+ int receiver = smpi_group_index(smpi_comm_group(request->comm), request->dst);
+/* if(receiver == MPI_UNDEFINED) {*/
+/* XBT_WARN("Trying to send a message to a wrong rank");*/
+/* return;*/
+/* }*/
+ print_request("New send", request);
+ if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
+ mailbox = smpi_process_remote_mailbox_small(receiver);
+ }else{
+ XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
+ mailbox = smpi_process_remote_mailbox(receiver);
+ }
+ if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
+ void *oldbuf = NULL;
+ request->detached = 1;
+ request->refcount++;
+ if(request->old_type->has_subtype == 0){
+ oldbuf = request->buf;
+ if (oldbuf){
+ request->buf = xbt_malloc(request->size);
+ memcpy(request->buf,oldbuf,request->size);
+ }
+ }
+ XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
+ }
+ // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
+ request->real_size=request->size;
+ smpi_datatype_use(request->old_type);
+
+ //if we are giving back the control to the user without waiting for completion, we have to inject timings
+ double sleeptime =0.0;
+ if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
+ //isend and send timings may be different
+ sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
+ }
+
+ if(sleeptime!=0.0){
+ simcall_process_sleep(sleeptime);
+ XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size));
+ }
+
+ request->action =
+ simcall_comm_isend(mailbox, request->size, -1.0,
+ request->buf, request->real_size,
+ &match_send,
+ &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails
+ request,
+ // detach if msg size < eager/rdv switch limit
+ request->detached);
+
+#ifdef HAVE_TRACING
+ /* FIXME: detached sends are not traceable (request->action == NULL) */
+ if (request->action)
+ simcall_set_category(request->action, TRACE_internal_smpi_get_category());
+#endif
+
+ }
}
-void smpi_global_destroy()
+void smpi_mpi_startall(int count, MPI_Request * requests)
{
- 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);
+ int i;
- // 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]);
- }
+ for(i = 0; i < count; i++) {
+ smpi_mpi_start(requests[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_mpi_request_free(MPI_Request * request)
+{
- xbt_free(smpi_global);
+ if((*request) != MPI_REQUEST_NULL){
+ (*request)->refcount--;
+ if((*request)->refcount<0) xbt_die("wrong refcount");
+
+ if((*request)->refcount==0){
+ xbt_free(*request);
+ *request = MPI_REQUEST_NULL;
+ }
+ }else{
+ xbt_die("freeing an already free request");
+ }
}
-int smpi_run_simulation(int argc, char **argv)
+MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype,
+ int dst, int tag, MPI_Comm comm)
{
- xbt_fifo_item_t cond_item = NULL;
- smx_cond_t cond = NULL;
- smx_action_t action = NULL;
+ MPI_Request request =
+ build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
+ comm, NON_PERSISTENT | SEND);
- xbt_fifo_t actions_failed = xbt_fifo_new();
- xbt_fifo_t actions_done = xbt_fifo_new();
+ return request;
+}
- srand(SMPI_RAND_SEED);
+MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype,
+ int dst, int tag, MPI_Comm comm)
+{
+ MPI_Request request =
+ build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
+ comm, NON_PERSISTENT | ISEND | SEND);
- SIMIX_global_init(&argc, argv);
+ smpi_mpi_start(request);
+ return request;
+}
- SIMIX_function_register("smpi_simulated_main", smpi_simulated_main);
- SIMIX_function_register("smpi_sender", smpi_sender);
- SIMIX_function_register("smpi_receiver", smpi_receiver);
+MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype,
+ int dst, int tag, MPI_Comm comm)
+{
+ MPI_Request request =
+ build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
+ comm, NON_PERSISTENT | ISEND | SSEND | SEND);
+ smpi_mpi_start(request);
+ return request;
+}
- // 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]);
+MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype,
+ int src, int tag, MPI_Comm comm)
+{
+ MPI_Request request =
+ build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
+ comm, NON_PERSISTENT | RECV);
+ return request;
+}
- /* Prepare to display some more info when dying on Ctrl-C pressing */
- // FIXME: doesn't work
- //signal(SIGINT, inthandler);
+MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
+ int src, int tag, MPI_Comm comm)
+{
+ MPI_Request request =
+ build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
+ comm, NON_PERSISTENT | RECV);
- /* Clean IO before the run */
- fflush(stdout);
- fflush(stderr);
+ smpi_mpi_start(request);
+ return request;
+}
- while (SIMIX_solve(actions_done, actions_failed) != -1.0) {
- while ((action = xbt_fifo_pop(actions_failed))) {
- DEBUG1("** %s failed **", action->name);
- xbt_fifo_foreach(action->cond_list, cond_item, cond, smx_cond_t) {
- SIMIX_cond_broadcast(cond);
- SIMIX_unregister_action_to_condition(action, cond);
- }
- SIMIX_action_destroy(action);
- }
- while ((action = xbt_fifo_pop(actions_done))) {
- DEBUG1("** %s done **",action->name);
- xbt_fifo_foreach(action->cond_list, cond_item, cond, smx_cond_t) {
- SIMIX_cond_broadcast(cond);
- SIMIX_unregister_action_to_condition(action, cond);
- }
- SIMIX_action_destroy(action);
- }
- }
+void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src,
+ int tag, MPI_Comm comm, MPI_Status * status)
+{
+ MPI_Request request;
+ request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
+ smpi_mpi_wait(&request, status);
+}
- xbt_fifo_free(actions_failed);
- xbt_fifo_free(actions_done);
- INFO1("simulation time %g", SIMIX_get_clock());
- smpi_global_destroy();
+void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
+ int tag, MPI_Comm comm)
+{
+ MPI_Request request =
+ build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
+ comm, NON_PERSISTENT | SEND);
- SIMIX_clean();
+ smpi_mpi_start(request);
+ smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
- return 0;
}
-void smpi_mpi_land_func(void *x, void *y, void *z)
+void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype,
+ int dst, int tag, MPI_Comm comm)
{
- *(int *)z = *(int *)x && *(int *)y;
+ MPI_Request request = smpi_mpi_issend(buf, count, datatype, dst, tag, comm);
+ smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
}
-void smpi_mpi_sum_func(void *x, void *y, void *z)
+void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
+ int dst, int sendtag, void *recvbuf, int recvcount,
+ MPI_Datatype recvtype, int src, int recvtag,
+ MPI_Comm comm, MPI_Status * status)
{
- *(int *)z = *(int *)x + *(int *)y;
+ MPI_Request requests[2];
+ MPI_Status stats[2];
+
+ requests[0] =
+ smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
+ requests[1] =
+ smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
+ smpi_mpi_startall(2, requests);
+ smpi_mpi_waitall(2, requests, stats);
+ if(status != MPI_STATUS_IGNORE) {
+ // Copy receive status
+ *status = stats[1];
+ }
}
-
-void smpi_mpi_init()
+int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
{
- 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->simdata = xbt_new(s_smpi_mpi_communicator_simdata_t, 1);
- smpi_mpi_global->mpi_comm_world->simdata->barrier_count = 0;
- smpi_mpi_global->mpi_comm_world->simdata->barrier_mutex = SIMIX_mutex_init();
- smpi_mpi_global->mpi_comm_world->simdata->barrier_cond = SIMIX_cond_init();
- smpi_mpi_global->mpi_comm_world->simdata->hosts = hosts;
- smpi_mpi_global->mpi_comm_world->simdata->processes = xbt_new(smx_process_t, size);
- smpi_mpi_global->mpi_comm_world->simdata->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;
- }
+ return status->count / smpi_datatype_size(datatype);
+}
- // 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);
+static void finish_wait(MPI_Request * request, MPI_Status * status)
+{
+ MPI_Request req = *request;
+ if(status != MPI_STATUS_IGNORE)
+ smpi_empty_status(status);
+
+ if(!(req->detached && req->flags & SEND)){
+ if(status != MPI_STATUS_IGNORE) {
+ status->MPI_SOURCE = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
+ status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag;
+ status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
+ // this handles the case were size in receive differs from size in send
+ // FIXME: really this should just contain the count of receive-type blocks,
+ // right?
+ status->count = req->real_size;
+ }
+
+ print_request("Finishing", req);
+ MPI_Datatype datatype = req->old_type;
+
+ if(datatype->has_subtype == 1){
+ // This part handles the problem of non-contignous memory
+ // the unserialization at the reception
+ s_smpi_subtype_t *subtype = datatype->substruct;
+ if(req->flags & RECV) {
+ subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct);
+ }
+ if(req->detached == 0) free(req->buf);
+ }
+ smpi_datatype_unuse(datatype);
+ }
+
+ if(req->detached_sender!=NULL){
+ smpi_mpi_request_free(&(req->detached_sender));
+ }
+
+ if(req->flags & NON_PERSISTENT) {
+ smpi_mpi_request_free(request);
+ } else {
+ req->action = NULL;
+ }
+}
- return;
+int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
+ int flag;
+
+ //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
+ if ((*request)->action == NULL)
+ flag = 1;
+ else
+ flag = simcall_comm_test((*request)->action);
+ if(flag) {
+ (*request)->refcount++;
+ finish_wait(request, status);
+ }else{
+ smpi_empty_status(status);
+ }
+ return flag;
}
-void smpi_mpi_finalize()
+int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
+ MPI_Status * status)
{
- int i;
+ xbt_dynar_t comms;
+ int i, flag, size;
+ int* map;
+
+ *index = MPI_UNDEFINED;
+ flag = 0;
+ if(count > 0) {
+ comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
+ map = xbt_new(int, count);
+ size = 0;
+ for(i = 0; i < count; i++) {
+ if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) {
+ xbt_dynar_push(comms, &requests[i]->action);
+ map[size] = i;
+ size++;
+ }
+ }
+ if(size > 0) {
+ i = simcall_comm_testany(comms);
+ // not MPI_UNDEFINED, as this is a simix return code
+ if(i != -1) {
+ *index = map[i];
+ finish_wait(&requests[*index], status);
+ flag = 1;
+ }
+ }else{
+ //all requests are null or inactive, return true
+ flag=1;
+ smpi_empty_status(status);
+ }
+ xbt_free(map);
+ xbt_dynar_free(&comms);
+ }
+
+ return flag;
+}
- 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);
+int smpi_mpi_testall(int count, MPI_Request requests[],
+ MPI_Status status[])
+{
+ MPI_Status stat;
+ MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
+ int flag=1;
+ int i;
+ for(i=0; i<count; i++){
+ if(requests[i]!= MPI_REQUEST_NULL){
+ if (smpi_mpi_test(&requests[i], pstat)!=1){
+ flag=0;
+ }
+ }else{
+ smpi_empty_status(pstat);
+ }
+ if(status != MPI_STATUSES_IGNORE) {
+ status[i] = *pstat;
+ }
+ }
+ return flag;
+}
- if (0 >= i) {
+void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
+ int flag=0;
+ //FIXME find another wait to avoid busy waiting ?
+ // the issue here is that we have to wait on a nonexistent comm
+ while(flag==0){
+ smpi_mpi_iprobe(source, tag, comm, &flag, status);
+ XBT_DEBUG("Busy Waiting on probing : %d", flag);
+ if(!flag) {
+ simcall_process_sleep(0.0001);
+ }
+ }
+}
- // 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]);
- }
- }
+void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
+ MPI_Request request =build_request(NULL, 0, MPI_CHAR, source, smpi_comm_rank(comm), tag,
+ comm, NON_PERSISTENT | RECV);
+
+ // behave like a receive, but don't do it
+ smx_rdv_t mailbox;
+
+ print_request("New iprobe", request);
+ // We have to test both mailboxes as we don't know if we will receive one one or another
+ if (sg_cfg_get_int("smpi/async_small_thres")>0){
+ mailbox = smpi_process_mailbox_small();
+ XBT_DEBUG("trying to probe the perm recv mailbox");
+ request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
+ }
+ if (request->action==NULL){
+ mailbox = smpi_process_mailbox();
+ XBT_DEBUG("trying to probe the other mailbox");
+ request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
+ }
+
+ if(request->action){
+ MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
+ *flag = 1;
+ if(status != MPI_STATUS_IGNORE) {
+ status->MPI_SOURCE = req->src;
+ status->MPI_TAG = req->tag;
+ status->MPI_ERROR = MPI_SUCCESS;
+ status->count = req->real_size;
+ }
+ }
+ else *flag = 0;
+ smpi_mpi_request_free(&request);
+
+ return;
+}
- // 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);
+void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
+{
+ print_request("Waiting", *request);
+ if ((*request)->action != NULL) { // this is not a detached send
+ simcall_comm_wait((*request)->action, -1.0);
+ }
+ finish_wait(request, status);
- 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);
+ // FIXME for a detached send, finish_wait is not called:
+}
- 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);
- }
-
-}
-
-// FIXME: could cause trouble with multithreaded procs on same host...
-void smpi_bench_begin()
+int smpi_mpi_waitany(int count, MPI_Request requests[],
+ MPI_Status * status)
{
- 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;
+ xbt_dynar_t comms;
+ int i, size, index;
+ int *map;
+
+ index = MPI_UNDEFINED;
+ if(count > 0) {
+ // Wait for a request to complete
+ comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
+ map = xbt_new(int, count);
+ size = 0;
+ XBT_DEBUG("Wait for one of %d", count);
+ for(i = 0; i < count; i++) {
+ if(requests[i] != MPI_REQUEST_NULL) {
+ if (requests[i]->action != NULL) {
+ XBT_DEBUG("Waiting any %p ", requests[i]);
+ xbt_dynar_push(comms, &requests[i]->action);
+ map[size] = i;
+ size++;
+ }else{
+ //This is a finished detached request, let's return this one
+ size=0;//so we free the dynar but don't do the waitany call
+ index=i;
+ finish_wait(&requests[i], status);//cleanup if refcount = 0
+ requests[i]=MPI_REQUEST_NULL;//set to null
+ break;
+ }
+ }
+ }
+ if(size > 0) {
+ i = simcall_comm_waitany(comms);
+
+ // not MPI_UNDEFINED, as this is a simix return code
+ if (i != -1) {
+ index = map[i];
+ finish_wait(&requests[index], status);
+ }
+ }
+ xbt_free(map);
+ xbt_dynar_free(&comms);
+ }
+
+ if (index==MPI_UNDEFINED)
+ smpi_empty_status(status);
+
+ return index;
}
-void smpi_bench_end()
+int smpi_mpi_waitall(int count, MPI_Request requests[],
+ MPI_Status status[])
{
- 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;
+ int index, c;
+ MPI_Status stat;
+ MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
+ int retvalue = MPI_SUCCESS;
+ //tag invalid requests in the set
+ if (status != MPI_STATUSES_IGNORE) {
+ for (c = 0; c < count; c++) {
+ if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL) {
+ smpi_empty_status(&status[c]);
+ } else if (requests[c]->src == MPI_PROC_NULL) {
+ smpi_empty_status(&status[c]);
+ status[c].MPI_SOURCE = MPI_PROC_NULL;
+ }
+ }
+ }
+ for(c = 0; c < count; c++) {
+ if(MC_is_active()) {
+ smpi_mpi_wait(&requests[c], pstat);
+ index = c;
+ } else {
+ index = smpi_mpi_waitany(count, requests, pstat);
+ if(index == MPI_UNDEFINED) {
+ break;
+ }
+ if (status != MPI_STATUSES_IGNORE) {
+ status[index] = *pstat;
+ if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
+ retvalue = MPI_ERR_IN_STATUS;
+ }
+ }
+ }
+
+ return retvalue;
}
-void smpi_barrier(smpi_mpi_communicator_t *comm)
+int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
+ MPI_Status status[])
{
-
- SIMIX_mutex_lock(comm->simdata->barrier_mutex);
- if(++comm->simdata->barrier_count < comm->size) {
- SIMIX_cond_wait(comm->simdata->barrier_cond, comm->simdata->barrier_mutex);
- } else {
- comm->simdata->barrier_count = 0;
- SIMIX_cond_broadcast(comm->simdata->barrier_cond);
- }
- SIMIX_mutex_unlock(comm->simdata->barrier_mutex);
-
- return;
+ int i, count, index;
+ MPI_Status stat;
+ MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
+
+ count = 0;
+ for(i = 0; i < incount; i++)
+ {
+ index=smpi_mpi_waitany(incount, requests, pstat);
+ if(index!=MPI_UNDEFINED){
+ indices[count] = index;
+ count++;
+ if(status != MPI_STATUSES_IGNORE) {
+ status[index] = *pstat;
+ }
+ }else{
+ return MPI_UNDEFINED;
+ }
+ }
+ return count;
}
-// FIXME: smarter algorithm...
-int smpi_comm_rank(smpi_mpi_communicator_t *comm, smx_host_t host)
+int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
+ MPI_Status status[])
{
- int i;
- for(i = 0; i < comm->size && host != comm->hosts[i]; i++);
- if (i >= comm->size) i = -1;
- return i;
+ int i, count, count_dead;
+ MPI_Status stat;
+ MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
+
+ count = 0;
+ count_dead = 0;
+ for(i = 0; i < incount; i++) {
+ if((requests[i] != MPI_REQUEST_NULL)) {
+ if(smpi_mpi_test(&requests[i], pstat)) {
+ indices[count] = i;
+ count++;
+ if(status != MPI_STATUSES_IGNORE) {
+ status[i] = *pstat;
+ }
+ }
+ }else{
+ count_dead++;
+ }
+ }
+ if(count_dead==incount)return MPI_UNDEFINED;
+ else return count;
}
-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)
+void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
+ MPI_Comm comm)
{
- 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;
+ // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
+ nary_tree_bcast(buf, count, datatype, root, comm, 4);
}
-int smpi_isend(smpi_mpi_request_t *request)
+void smpi_mpi_barrier(MPI_Comm comm)
{
- 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]);
- }
-
- if (SIMIX_process_is_suspended(smpi_global->sender_processes[rank])) {
- SIMIX_process_resume(smpi_global->sender_processes[rank]);
- }
-
- return retval;
+ // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
+ nary_tree_barrier(comm, 4);
}
-int smpi_irecv(smpi_mpi_request_t *request)
+void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
+ void *recvbuf, int recvcount, MPI_Datatype recvtype,
+ int root, MPI_Comm comm)
{
- 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_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]);
- }
+ int system_tag = 666;
+ int rank, size, src, index;
+ MPI_Aint lb = 0, recvext = 0;
+ MPI_Request *requests;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ if(rank != root) {
+ // Send buffer to root
+ smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
+ } else {
+ // FIXME: check for errors
+ smpi_datatype_extent(recvtype, &lb, &recvext);
+ // Local copy from root
+ smpi_datatype_copy(sendbuf, sendcount, sendtype,
+ (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
+ // Receive buffers from senders
+ requests = xbt_new(MPI_Request, size - 1);
+ index = 0;
+ for(src = 0; src < size; src++) {
+ if(src != root) {
+ requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
+ recvcount, recvtype,
+ src, system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of irecv's.
+ smpi_mpi_startall(size - 1, requests);
+ smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
+ xbt_free(requests);
+ }
+}
- if (SIMIX_process_is_suspended(smpi_global->receiver_processes[rank])) {
- SIMIX_process_resume(smpi_global->receiver_processes[rank]);
- }
+void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
+ void *recvbuf, int *recvcounts, int *displs,
+ MPI_Datatype recvtype, int root, MPI_Comm comm)
+{
+ int system_tag = 666;
+ int rank, size, src, index;
+ MPI_Aint lb = 0, recvext = 0;
+ MPI_Request *requests;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ if(rank != root) {
+ // Send buffer to root
+ smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
+ } else {
+ // FIXME: check for errors
+ smpi_datatype_extent(recvtype, &lb, &recvext);
+ // Local copy from root
+ smpi_datatype_copy(sendbuf, sendcount, sendtype,
+ (char *)recvbuf + displs[root] * recvext,
+ recvcounts[root], recvtype);
+ // Receive buffers from senders
+ requests = xbt_new(MPI_Request, size - 1);
+ index = 0;
+ for(src = 0; src < size; src++) {
+ if(src != root) {
+ requests[index] =
+ smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
+ recvcounts[src], recvtype, src, system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of irecv's.
+ smpi_mpi_startall(size - 1, requests);
+ smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
+ xbt_free(requests);
+ }
+}
- return retval;
+void smpi_mpi_allgather(void *sendbuf, int sendcount,
+ MPI_Datatype sendtype, void *recvbuf,
+ int recvcount, MPI_Datatype recvtype,
+ MPI_Comm comm)
+{
+ int system_tag = 666;
+ int rank, size, other, index;
+ MPI_Aint lb = 0, recvext = 0;
+ MPI_Request *requests;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ // FIXME: check for errors
+ smpi_datatype_extent(recvtype, &lb, &recvext);
+ // Local copy from self
+ smpi_datatype_copy(sendbuf, sendcount, sendtype,
+ (char *)recvbuf + rank * recvcount * recvext, recvcount,
+ recvtype);
+ // Send/Recv buffers to/from others;
+ requests = xbt_new(MPI_Request, 2 * (size - 1));
+ index = 0;
+ for(other = 0; other < size; other++) {
+ if(other != rank) {
+ requests[index] =
+ smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
+ comm);
+ index++;
+ requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
+ recvcount, recvtype, other,
+ system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of all comms.
+ smpi_mpi_startall(2 * (size - 1), requests);
+ smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
+ xbt_free(requests);
}
-void smpi_wait(smpi_mpi_request_t *request, smpi_mpi_status_t *status)
+void smpi_mpi_allgatherv(void *sendbuf, int sendcount,
+ MPI_Datatype sendtype, void *recvbuf,
+ int *recvcounts, int *displs,
+ MPI_Datatype recvtype, MPI_Comm comm)
{
- if (NULL != request) {
- SIMIX_mutex_lock(request->simdata->mutex);
- if (!request->completed) {
- SIMIX_cond_wait(request->simdata->cond, request->simdata->mutex);
- }
- if (NULL != status) {
- status->MPI_SOURCE = request->src;
- }
- SIMIX_mutex_unlock(request->simdata->mutex);
- }
+ int system_tag = 666;
+ int rank, size, other, index;
+ MPI_Aint lb = 0, recvext = 0;
+ MPI_Request *requests;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ // FIXME: check for errors
+ smpi_datatype_extent(recvtype, &lb, &recvext);
+ // Local copy from self
+ smpi_datatype_copy(sendbuf, sendcount, sendtype,
+ (char *)recvbuf + displs[rank] * recvext,
+ recvcounts[rank], recvtype);
+ // Send buffers to others;
+ requests = xbt_new(MPI_Request, 2 * (size - 1));
+ index = 0;
+ for(other = 0; other < size; other++) {
+ if(other != rank) {
+ requests[index] =
+ smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
+ comm);
+ index++;
+ requests[index] =
+ smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
+ recvtype, other, system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of all comms.
+ smpi_mpi_startall(2 * (size - 1), requests);
+ smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
+ xbt_free(requests);
}
-// FIXME: move into own file
-int smpi_gettimeofday(struct timeval *tv, struct timezone *tz)
+void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
+ void *recvbuf, int recvcount, MPI_Datatype recvtype,
+ int root, MPI_Comm comm)
{
- 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);
- }
- smpi_bench_begin();
- return retval;
-}
-
-unsigned int smpi_sleep(unsigned int seconds)
-{
- smx_mutex_t mutex;
- smx_cond_t cond;
- smx_host_t host;
- smx_action_t sleep_action;
-
- smpi_bench_end();
- host = SIMIX_host_self();
- sleep_action = SIMIX_action_sleep(host, seconds);
- mutex = SIMIX_mutex_init();
- cond = SIMIX_cond_init();
-
- SIMIX_mutex_lock(mutex);
- SIMIX_register_action_to_condition(sleep_action, cond);
- SIMIX_cond_wait(cond, mutex);
- //SIMIX_unregister_action_to_condition(sleep_action, cond);
- SIMIX_mutex_unlock(mutex);
+ int system_tag = 666;
+ int rank, size, dst, index;
+ MPI_Aint lb = 0, sendext = 0;
+ MPI_Request *requests;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ if(rank != root) {
+ // Recv buffer from root
+ smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
+ MPI_STATUS_IGNORE);
+ } else {
+ // FIXME: check for errors
+ smpi_datatype_extent(sendtype, &lb, &sendext);
+ // Local copy from root
+ smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
+ sendcount, sendtype, recvbuf, recvcount, recvtype);
+ // Send buffers to receivers
+ requests = xbt_new(MPI_Request, size - 1);
+ index = 0;
+ for(dst = 0; dst < size; dst++) {
+ if(dst != root) {
+ requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
+ sendcount, sendtype, dst,
+ system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of isend's.
+ smpi_mpi_startall(size - 1, requests);
+ smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
+ xbt_free(requests);
+ }
+}
- SIMIX_mutex_destroy(mutex);
- SIMIX_cond_destroy(cond);
+void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
+ MPI_Datatype sendtype, void *recvbuf, int recvcount,
+ MPI_Datatype recvtype, int root, MPI_Comm comm)
+{
+ int system_tag = 666;
+ int rank, size, dst, index;
+ MPI_Aint lb = 0, sendext = 0;
+ MPI_Request *requests;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ if(rank != root) {
+ // Recv buffer from root
+ smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
+ MPI_STATUS_IGNORE);
+ } else {
+ // FIXME: check for errors
+ smpi_datatype_extent(sendtype, &lb, &sendext);
+ // Local copy from root
+ smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
+ sendtype, recvbuf, recvcount, recvtype);
+ // Send buffers to receivers
+ requests = xbt_new(MPI_Request, size - 1);
+ index = 0;
+ for(dst = 0; dst < size; dst++) {
+ if(dst != root) {
+ requests[index] =
+ smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
+ sendtype, dst, system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of isend's.
+ smpi_mpi_startall(size - 1, requests);
+ smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
+ xbt_free(requests);
+ }
+}
- // FIXME: check for success/failure?
+void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
+ MPI_Datatype datatype, MPI_Op op, int root,
+ MPI_Comm comm)
+{
+ int system_tag = 666;
+ int rank, size, src, index;
+ MPI_Aint lb = 0, dataext = 0;
+ MPI_Request *requests;
+ void **tmpbufs;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ if(rank != root) {
+ // Send buffer to root
+ smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm);
+ } else {
+ // FIXME: check for errors
+ smpi_datatype_extent(datatype, &lb, &dataext);
+ // Local copy from root
+ if (sendbuf && recvbuf)
+ smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
+ // Receive buffers from senders
+ //TODO: make a MPI_barrier here ?
+ requests = xbt_new(MPI_Request, size - 1);
+ tmpbufs = xbt_new(void *, size - 1);
+ index = 0;
+ for(src = 0; src < size; src++) {
+ if(src != root) {
+ // FIXME: possibly overkill we we have contiguous/noncontiguous data
+ // mapping...
+ tmpbufs[index] = xbt_malloc(count * dataext);
+ requests[index] =
+ smpi_irecv_init(tmpbufs[index], count, datatype, src,
+ system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of irecv's.
+ smpi_mpi_startall(size - 1, requests);
+ for(src = 0; src < size - 1; src++) {
+ index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
+ XBT_DEBUG("finished waiting any request with index %d", index);
+ if(index == MPI_UNDEFINED) {
+ break;
+ }
+ if(op) /* op can be MPI_OP_NULL that does nothing */
+ smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+ }
+ for(index = 0; index < size - 1; index++) {
+ xbt_free(tmpbufs[index]);
+ }
+ xbt_free(tmpbufs);
+ xbt_free(requests);
+ }
+}
- smpi_bench_begin();
- return 0;
+void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
+ MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
+{
+ smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
+ smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
}
-void smpi_exit(int status)
+void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
+ MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
{
- 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;
+ int system_tag = 666;
+ int rank, size, other, index;
+ MPI_Aint lb = 0, dataext = 0;
+ MPI_Request *requests;
+ void **tmpbufs;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+
+ // FIXME: check for errors
+ smpi_datatype_extent(datatype, &lb, &dataext);
+
+ // Local copy from self
+ smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
+
+ // Send/Recv buffers to/from others;
+ requests = xbt_new(MPI_Request, size - 1);
+ tmpbufs = xbt_new(void *, rank);
+ index = 0;
+ for(other = 0; other < rank; other++) {
+ // FIXME: possibly overkill we we have contiguous/noncontiguous data
+ // mapping...
+ tmpbufs[index] = xbt_malloc(count * dataext);
+ requests[index] =
+ smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
+ comm);
+ index++;
+ }
+ for(other = rank + 1; other < size; other++) {
+ requests[index] =
+ smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
+ index++;
+ }
+ // Wait for completion of all comms.
+ smpi_mpi_startall(size - 1, requests);
+ for(other = 0; other < size - 1; other++) {
+ index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
+ if(index == MPI_UNDEFINED) {
+ break;
+ }
+ if(index < rank) {
+ // #Request is below rank: it's a irecv
+ smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+ }
+ }
+ for(index = 0; index < rank; index++) {
+ xbt_free(tmpbufs[index]);
+ }
+ xbt_free(tmpbufs);
+ xbt_free(requests);
}