#include <signal.h>
#include <sys/time.h>
-#include "xbt/xbt_portability.h"
-#include "simix/simix.h"
-#include "simix/private.h"
+#include "xbt/xbt_os_time.h"
+#include "xbt/mallocator.h"
#include "smpi.h"
// FIXME: move globals into structure...
-xbt_mallocator_t smpi_request_mallocator = NULL;
-xbt_mallocator_t smpi_message_mallocator = NULL;
-xbt_fifo_t *smpi_pending_send_requests = NULL;
-xbt_fifo_t *smpi_pending_recv_requests = NULL;
-xbt_fifo_t *smpi_received_messages = NULL;
+xbt_mallocator_t smpi_request_mallocator = NULL;
+xbt_mallocator_t smpi_message_mallocator = NULL;
+
+xbt_fifo_t *smpi_pending_send_requests = NULL;
+smx_mutex_t *smpi_pending_send_requests_mutex = NULL;
+
+xbt_fifo_t *smpi_pending_recv_requests = NULL;
+smx_mutex_t *smpi_pending_recv_requests_mutex = NULL;
+
+xbt_fifo_t *smpi_received_messages = NULL;
+smx_mutex_t *smpi_received_messages_mutex = NULL;
smx_process_t *smpi_sender_processes = NULL;
smx_process_t *smpi_receiver_processes = NULL;
smx_mutex_t smpi_running_hosts_mutex = NULL;
smx_mutex_t smpi_benchmarking_mutex = NULL;
smx_mutex_t init_mutex = NULL;
-smx_cond_t init_cond = NULL;
+smx_cond_t init_cond = NULL;
-int smpi_root_ready = 0;
+int smpi_root_ready = 0;
int smpi_ready_count = 0;
XBT_LOG_NEW_DEFAULT_CATEGORY(smpi, "SMPI");
int inline smpi_mpi_comm_world_rank_self()
{
- return smpi_mpi_comm_rank(&smpi_mpi_comm_world, SIMIX_host_self())
+ return smpi_mpi_comm_rank(&smpi_mpi_comm_world, SIMIX_host_self());
}
-// FIXME: messages are actually smaller than requests, use them instead?
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 = 0;
+
smpi_mpi_request_t *request;
+
smx_host_t dhost;
+
smx_action_t communicate_action;
- smpi_mpi_request_t *scratch;
+
+ smpi_received_message_t *message;
+
int drank;
+
smx_process_t waitproc;
self = SIMIX_process_self();
}
SIMIX_mutex_unlock(init_mutex);
- request_queue = smpi_pending_send_requests[rank];
- size = smpi_mpi_comm_size(&smpi_mpi_comm_world);
+ request_queue = smpi_pending_send_requests[rank];
+ request_queue_mutex = smpi_pending_send_requests_mutex[rank];
+ size = smpi_mpi_comm_size(&smpi_mpi_comm_world);
+
smpi_sender_processes[rank] = self;
// wait for all nodes to signal initializatin complete
while (0 < running_hosts) {
- // FIXME: mutex?
+ 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);
SIMIX_cond_wait(request->cond, request->mutex);
// copy request to appropriate received queue
- scratch = xbt_mallocator_get(smpi_message_mallocator);
- scratch->comm = request->comm;
- scratch->src = request->src;
- scratch->dst = request->dst;
- scratch->tag = request->tag;
- scratch->buf = request->buf;
+ message = xbt_mallocator_get(smpi_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);
+
drank = smpi_mpi_comm_rank(&smpi_mpi_comm_world, dhost);
- xbt_fifo_push(smpi_received_messages[drank], scratch);
+
+ SIMIX_mutex_lock(smpi_received_messages_mutex[drank]);
+ xbt_fifo_push(smpi_received_messages[drank], message);
+ SIMIX_mutex_unlock(smpi_received_messages_mutex[drank]);
request->completed = 1;
- while(waitproc = xbt_fifo_shift(request->waitlist)) {
+ // wake up receiver, then any waiting sender
+ waitproc = smpi_receiver_processes[drank];
+
+ do {
if (SIMIX_process_is_suspended(waitproc)) {
SIMIX_process_resume(waitproc);
}
- }
+ } while(waitproc = xbt_fifo_shift(request->waitlist));
SIMIX_mutex_unlock(request->mutex);
}
SIMIX_mutex_unlock(smpi_running_hosts_mutex);
}
+ SIMIX_mutex_lock(init_mutex);
+ smpi_ready_count--;
+ if (smpi_ready_count <= 0) {
+ SIMIX_cond_broadcast(init_cond);
+ }
+ SIMIX_mutex_unlock(init_mutex);
+
return 0;
}
{
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;
- smpi_mpi_request_t *message;
+
smpi_mpi_request_t *request;
+ smpi_received_message_t *message;
+
+ xbt_fifo_item_t request_item;
+ xbt_fifo_item_t message_item;
+
smx_process_t waitproc;
self = SIMIX_process_self();
}
SIMIX_mutex_unlock(init_mutex);
- request_queue = smpi_pending_receive_requests[rank];
- message_queue = smpi_received_messages[rank];
- size = smpi_mpi_comm_size(&smpi_mpi_comm_world);
+ request_queue = smpi_pending_recv_requests[rank];
+ request_queue_mutex = smpi_pending_recv_requests_mutex[rank];
+ message_queue = smpi_received_messages[rank];
+ message_queue_mutex = smpi_received_messages_mutex[rank];
+ size = smpi_mpi_comm_size(&smpi_mpi_comm_world);
+
smpi_receiver_processes[rank] = self;
// wait for all nodes to signal initializatin complete
while (0 < running_hosts) {
- // FIXME: search for received messages and requests
- // use stupid algorithm for now
+ 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) {
+ if (NULL == request || NULL == message) {
SIMIX_process_suspend(self);
} else {
SIMIX_mutex_lock(request->mutex);
- memcpy(request->buf, message->buf, request->count * request->type->size);
+
+ memcpy(request->buf, message->buf, request->count * request->datatype->size);
request->src = message->src;
- reqeust->completed = 1;
+ request->completed = 1;
while (waitproc = xbt_fifo_shift(request->waitlist)) {
if (SIMIX_process_is_suspended(waitproc)) {
}
SIMIX_mutex_unlock(request->mutex);
+
+ xbt_free(message->buf);
xbt_mallocator_release(smpi_message_mallocator, message);
}
SIMIX_mutex_unlock(smpi_running_hosts_mutex);
}
+ SIMIX_mutex_lock(init_mutex);
+ smpi_ready_count--;
+ if (smpi_ready_count <= 0) {
+ SIMIX_cond_broadcast(init_cond);
+ }
+ SIMIX_mutex_unlock(init_mutex);
+
return 0;
}
SIMIX_global_init(&argc, argv);
+ // important globals
init_mutex = SIMIX_mutex_init();
init_cond = SIMIX_cond_init();
SIMIX_function_register("smpi_simulated_main", smpi_simulated_main);
- SIMIX_function_register("smpi_sender", smpi_sender);
- SIMIX_function_register("smpi_receiver", smpi_receiver);
+ 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]);
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 */
SIMIX_action_destroy(action);
}
}
+
xbt_fifo_free(actions_failed);
xbt_fifo_free(actions_done);
+
INFO1("simulation time %g", SIMIX_get_clock());
+
SIMIX_clean();
+
return 0;
}
*(int *)z = *(int *)x + *(int *)y;
}
-smpi_mpi_request_t *smpi_new_request()
+void *smpi_new_request()
{
return xbt_new(smpi_mpi_request_t, 1);
}
+void smpi_free_request(void *pointer) {
+ smpi_mpi_request_t *request = pointer;
+
+ if (NULL != request) {
+ SIMIX_mutex_destroy(request->mutex);
+ SIMIX_cond_destroy(request->cond);
+ xbt_fifo_free(request->waitlist);
+ xbt_free(request);
+ }
+
+ return;
+}
+
+void smpi_reset_request(void *pointer) {
+ smpi_mpi_request_t *request = pointer;
+
+ if (NULL != request) {
+ request->mutex = SIMIX_mutex_init();
+ request->cond = SIMIX_cond_init();
+ request->waitlist = xbt_fifo_new();
+ // FIXME: clear waitlist
+ }
+
+ return;
+}
+
+
+void *smpi_new_message()
+{
+ return xbt_new(smpi_received_message_t, 1);
+}
+
+void smpi_do_nothing(void *pointer)
+{
+ return;
+}
+
void smpi_mpi_init()
{
int i;
smpi_mpi_sum.func = &smpi_mpi_sum_func;
// smpi globals
- smpi_request_mallocator = xbt_mallocator_new(SMPI_REQUEST_MALLOCATOR_SIZE, smpi_new_request, xbt_free, NULL);
- smpi_message_mallocator = xbt_mallocator_new(SMPI_MESSAGE_MALLOCATOR_SIZE, smpi_new_message, xbt_free, NULL);
- smpi_pending_send_requests = xbt_new(xbt_fifo_t, size);
- smpi_pending_recv_requests = xbt_new(xbt_fifo_t, size);
- smpi_received_messages = xbt_new(xbt_fifo_t, size);
+ smpi_request_mallocator = xbt_mallocator_new(SMPI_REQUEST_MALLOCATOR_SIZE, smpi_new_request, smpi_free_request, smpi_reset_request);
+ smpi_message_mallocator = xbt_mallocator_new(SMPI_MESSAGE_MALLOCATOR_SIZE, smpi_new_message, smpi_free_message, smpi_do_nothing);
+ smpi_pending_send_requests = xbt_new(xbt_fifo_t, size);
+ smpi_pending_send_requests_mutex = xbt_new(smx_mutex_t, size);
+ smpi_pending_recv_requests = xbt_new(xbt_fifo_t, size);
+ smpi_pending_recv_requests_mutex = xbt_new(smx_mutex_t, size);
+ smpi_received_messages = xbt_new(xbt_fifo_t, size);
+ smpi_received_messages_mutex = xbt_new(smx_mutex_t, size);
for(i = 0; i < size; i++) {
- smpi_pending_send_requests[i] = xbt_fifo_new();
- smpi_pending_recv_requests[i] = xbt_fifo_new();
- smpi_received_messages[i] = xbt_fifo_new();
+ smpi_pending_send_requests[i] = xbt_fifo_new();
+ smpi_pending_send_requests_mutex[i] = SIMIX_mutex_init();
+ smpi_pending_recv_requests[i] = xbt_fifo_new();
+ smpi_pending_recv_requests_mutex[i] = SIMIX_mutex_init();
+ smpi_received_messages[i] = xbt_fifo_new();
+ smpi_received_messages_mutex[i] = SIMIX_mutex_init();
}
smpi_timer = xbt_os_timer_new();
i = --smpi_running_hosts;
SIMIX_mutex_unlock(smpi_running_hosts_mutex);
+ SIMIX_mutex_lock(init_mutex);
+ smpi_ready_count--;
+ SIMIX_mutex_unlock(init_mutex);
+
if (0 >= i) {
+ // wake up senders/receivers
+ for (i = 0; i < smpi_mpi_comm_world.size; i++) {
+ if (SIMIX_process_is_suspended(smpi_sender_processes[i])) {
+ SIMIX_process_resume(smpi_sender_processes[i]);
+ }
+ if (SIMIX_process_is_suspended(smpi_receiver_processes[i])) {
+ SIMIX_process_resume(smpi_receiver_processes[i]);
+ }
+ }
+
+ // wait for senders/receivers to exit...
+ SIMIX_mutex_lock(init_mutex);
+ if (smpi_ready_count > 0) {
+ SIMIX_cond_wait(init_cond, init_mutex);
+ }
+ SIMIX_mutex_unlock(init_mutex);
+
+ SIMIX_mutex_destroy(init_mutex);
+ SIMIX_cond_destroy(init_cond);
SIMIX_mutex_destroy(smpi_running_hosts_mutex);
for (i = 0 ; i < smpi_mpi_comm_world.size; i++) {
xbt_fifo_free(smpi_pending_send_requests[i]);
+ SIMIX_mutex_destroy(smpi_pending_send_requests_mutex[i]);
xbt_fifo_free(smpi_pending_recv_requests[i]);
+ SIMIX_mutex_destroy(smpi_pending_recv_requests_mutex[i]);
xbt_fifo_free(smpi_received_messages[i]);
+ SIMIX_mutex_destroy(smpi_received_messages_mutex[i]);
}
xbt_mallocator_free(smpi_request_mallocator);
xbt_mallocator_free(smpi_message_mallocator);
xbt_free(smpi_pending_send_requests);
+ xbt_free(smpi_pending_send_requests_mutex);
xbt_free(smpi_pending_recv_requests);
+ xbt_free(smpi_pending_recv_requests_mutex);
xbt_free(smpi_received_messages);
+ xbt_free(smpi_received_messages_mutex);
SIMIX_mutex_destroy(smpi_mpi_comm_world.barrier_mutex);
SIMIX_cond_destroy(smpi_mpi_comm_world.barrier_cond);
void smpi_bench_begin()
{
+ SIMIX_mutex_lock(smpi_benchmarking_mutex);
xbt_assert0(!smpi_benchmarking, "Already benchmarking");
smpi_benchmarking = 1;
+ SIMIX_mutex_unlock(smpi_benchmarking_mutex);
+
xbt_os_timer_start(smpi_timer);
+
return;
}
smx_mutex_t mutex;
smx_cond_t cond;
+ SIMIX_mutex_lock(smpi_benchmarking_mutex);
xbt_assert0(smpi_benchmarking, "Not benchmarking yet");
smpi_benchmarking = 0;
+ SIMIX_mutex_lock(smpi_benchmarking_mutex);
+
xbt_os_timer_stop(smpi_timer);
- duration = xbt_os_timer_elapsed(smpi_timer);
+
+ duration = xbt_os_timer_elapsed(smpi_timer);
host = SIMIX_host_self();
compute_action = SIMIX_action_execute(host, "computation", duration * SMPI_DEFAULT_SPEED);
mutex = SIMIX_mutex_init();
cond = SIMIX_cond_init();
+
SIMIX_mutex_lock(mutex);
SIMIX_register_condition_to_action(compute_action, cond);
SIMIX_register_action_to_condition(compute_action, cond);
SIMIX_mutex_unlock(mutex);
SIMIX_mutex_destroy(mutex);
SIMIX_cond_destroy(cond);
+
// FIXME: check for success/failure?
+
return;
}
void smpi_barrier(smpi_mpi_communicator_t *comm) {
int i;
+
SIMIX_mutex_lock(comm->barrier_mutex);
comm->barrier++;
if(i < comm->size) {
SIMIX_cond_broadcast(comm->barrier_cond);
}
SIMIX_mutex_unlock(comm->barrier_mutex);
+
+ return;
}
+// FIXME: smarter algorithm...
int smpi_comm_rank(smpi_mpi_communicator_t *comm, smx_host_t host)
{
int i;
retval = MPI_ERR_TAG;
} else {
*request = xbt_mallocator_get(smpi_request_mallocator);
- (*request)->buf = buf;
- (*request)->count = count;
- (*request)->datatype = datatype;
+ (*request)->comm = comm;
(*request)->src = src;
(*request)->dst = dst;
(*request)->tag = tag;
- (*request)->comm = comm;
+ (*request)->buf = buf;
+ (*request)->count = count;
+ (*request)->datatype = datatype;
(*request)->completed = 0;
- (*request)->waitlist = NULL;
+ (*request)->mutex = SIMIX_mutex_init();
+ (*request)->cond = SIMIX_cond_init();
+ (*request)->waitlist = xbt_fifo_new();
}
return retval;
}
int smpi_isend(smpi_mpi_request_t *request)
{
- int rank = smpi_mpi_comm_rank_self(&smpi_mpi_comm_world);
+ int retval = MPI_SUCCESS;
+ int rank = smpi_mpi_comm_rank_self(&smpi_mpi_comm_world);
- xbt_fifo_push(smpi_pending_send_requests[rank], request);
+ if (NULL != request) {
+ SIMIX_mutex_lock(smpi_pending_send_requests_mutex[rank]);
+ xbt_fifo_push(smpi_pending_send_requests[rank], request);
+ SIMIX_mutex_unlock(smpi_pending_send_requests_mutex[rank]);
+ }
- if (MSG_process_is_suspended(smpi_sender_processes[rank])) {
- MSG_process_resume(smpi_sender_processes[rank]);
+ if (SIMIX_process_is_suspended(smpi_sender_processes[rank])) {
+ SIMIX_process_resume(smpi_sender_processes[rank]);
}
+
+ return retval;
}
int smpi_irecv(smpi_mpi_request_t *request)
{
+ int retval = MPI_SUCCESS;
int rank = smpi_mpi_comm_rank_self(&smpi_mpi_comm_world);
- xbt_fifo_push(smpi_pending_recv_requests[rank], request);
+ if (NULL != request) {
+ SIMIX_mutex_lock(smpi_pending_recv_requests_mutex[rank]);
+ xbt_fifo_push(smpi_pending_recv_requests[rank], request);
+ SIMIX_mutex_unlock(smpi_pending_recv_requests_mutex[rank]);
+ }
- if (MSG_process_is_suspended(smpi_receiver_processes[rank])) {
- MSG_process_resume(smpi_receiver_processes[rank]);
+ if (SIMIX_process_is_suspended(smpi_receiver_processes[rank])) {
+ SIMIX_process_resume(smpi_receiver_processes[rank]);
}
+
+ return retval;
}
void smpi_wait(smpi_mpi_request_t *request, smpi_mpi_status_t *status)
{
- smx_process_t self;
+ smx_process_t self = SIMIX_process_self();
int suspend = 0;
- self = SIMIX_process_self();
if (NULL != request) {
SIMIX_mutex_lock(request->mutex);
}
SIMIX_mutex_unlock(request->mutex);
if (suspend) {
- SIMIX_suspend(self);
+ SIMIX_process_suspend(self);
}
if (NULL != status && MPI_STATUS_IGNORE != status) {
SIMIX_mutex_lock(request->mutex);