#include <stdio.h>
+
+#include <signal.h>
#include <sys/time.h>
-#include "msg/msg.h"
-#include "simix/simix.h"
-#include "xbt/sysdep.h"
-#include "xbt/xbt_portability.h"
+#include "xbt/xbt_os_time.h"
+#include "xbt/mallocator.h"
#include "smpi.h"
-smpi_mpi_request_t **smpi_pending_send_requests = NULL;
-smpi_mpi_request_t **smpi_last_pending_send_requests = NULL;
+// FIXME: move globals into structure...
+
+xbt_mallocator_t smpi_request_mallocator = NULL;
+xbt_mallocator_t smpi_message_mallocator = NULL;
-smpi_mpi_request_t **smpi_pending_recv_requests = NULL;
-smpi_mpi_request_t **smpi_last_pending_recv_requests = NULL;
+xbt_fifo_t *smpi_pending_send_requests = NULL;
+smx_mutex_t *smpi_pending_send_requests_mutex = NULL;
-smpi_received_t **smpi_received = NULL;
-smpi_received_t **smpi_last_received = NULL;
+xbt_fifo_t *smpi_pending_recv_requests = NULL;
+smx_mutex_t *smpi_pending_recv_requests_mutex = NULL;
-smx_process_t *smpi_sender_processes = NULL;
-smx_process_t *smpi_receiver_processes = 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;
int smpi_running_hosts = 0;
static xbt_os_timer_t smpi_timer;
static int smpi_benchmarking;
-static double smpi_reference;
-
-int smpi_run_simulation(int *argc, char **argv) {
- smx_cond_t cond = NULL;
- smx_action_t smx_action;
- xbt_fifo_t actions_done = xbt_fifo_new();
- xbt_fifo_t actions_failed = xbt_fifo_new();
-
- srand(SEED);
- SIMIX_global_init(&argc, argv);
- SIMIX_function_register("smpi_main", smpi_main);
- SIMIX_function_register("smpi_sender", smpi_sender);
- SIMIX_function_register("smpi_receiver", smpi_receiver);
- SIMIX_create_environment(argv[1]);
- SIMIX_launch_application(argv[2]);
-
- /* Prepare to display some more info when dying on Ctrl-C pressing */
- signal(SIGINT,inthandler);
-
- /* Clean IO before the run */
- fflush(stdout);
- fflush(stderr);
-
- //surf_solve(); /* Takes traces into account. Returns 0.0 */
- /* xbt_fifo_size(msg_global->process_to_run) */
-
- while (SIMIX_solve(actions_done, actions_failed) != -1.0) {
-
- while ( (smx_action = xbt_fifo_pop(actions_failed)) ) {
-
-
- DEBUG1("** %s failed **",smx_action->name);
- while ( (cond = xbt_fifo_pop(smx_action->cond_list)) ) {
- SIMIX_cond_broadcast(cond);
- }
- /* action finished, destroy it */
- // SIMIX_action_destroy(smx_action);
- }
-
- while ( (smx_action = xbt_fifo_pop(actions_done)) ) {
-
- DEBUG1("** %s done **",smx_action->name);
- while ( (cond = xbt_fifo_pop(smx_action->cond_list)) ) {
- SIMIX_cond_broadcast(cond);
- }
- /* action finished, destroy it */
- //SIMIX_action_destroy(smx_action);
- }
- }
- xbt_fifo_free(actions_failed);
- xbt_fifo_free(actions_done);
- INFO1("simulation time %g", SIMIX_get_clock());
- SIMIX_clean();
- return 0;
-}
+static double smpi_reference_speed;
-void smpi_mpi_land_func(void *x, void *y, void *z) {
- *(int *)z = *(int *)x && *(int *)y;
-}
+// mutexes
+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;
-void smpi_mpi_sum_func(void *x, void *y, void *z) {
- *(int *)z = *(int *)x + *(int *)y;
-}
+int smpi_root_ready = 0;
+int smpi_ready_count = 0;
-void smpi_mpi_init() {
- int i;
- int size, rank;
- smx_host_t *hosts;
- smx_host_t host;
- double duration;
- m_task_t mtask;
-
- // will eventually need mutex
- smpi_running_hosts++;
-
- // initialize some local variables
- size = SIMIX_host_get_number();
- host = SIMIX_host_self();
- hosts = SIMIX_host_get_table();
- for(i = 0; i < size && host != hosts[i]; i++);
- rank = i;
-
- // node 0 sets the globals
- if (0 == rank) {
-
- // global communicator
- smpi_mpi_comm_world.id = 0;
- smpi_mpi_comm_world.size = size;
- smpi_mpi_comm_world.barrier = 0;
- smpi_mpi_comm_world.hosts = hosts;
- smpi_mpi_comm_world.processes = xbt_malloc(sizeof(m_process_t) * size);
- smpi_mpi_comm_world.processes[0] = SIMIX_process_self();
-
- // mpi datatypes
- smpi_mpi_byte.size = (size_t)1;
- smpi_mpi_int.size = sizeof(int);
- smpi_mpi_double.size = sizeof(double);
-
- // mpi operations
- smpi_mpi_land.func = &smpi_mpi_land_func;
- smpi_mpi_sum.func = &smpi_mpi_sum_func;
-
- // smpi globals
- smpi_pending_send_requests = xbt_malloc(sizeof(smpi_mpi_request_t*) * size);
- smpi_last_pending_send_requests = xbt_malloc(sizeof(smpi_mpi_request_t*) * size);
- smpi_pending_recv_requests = xbt_malloc(sizeof(smpi_mpi_request_t*) * size);
- smpi_last_pending_recv_requests = xbt_malloc(sizeof(smpi_mpi_request_t*) * size);
- smpi_received = xbt_malloc(sizeof(smpi_received_t*) * size);
- smpi_last_received = xbt_malloc(sizeof(smpi_received_t*) * size);
- smpi_sender_processes = xbt_malloc(sizeof(m_process_t) * size);
- smpi_receiver_processes = xbt_malloc(sizeof(m_process_t) * size);
- for(i = 0; i < size; i++) {
- smpi_pending_send_requests[i] = NULL;
- smpi_last_pending_send_requests[i] = NULL;
- smpi_pending_recv_requests[i] = NULL;
- smpi_last_pending_recv_requests[i] = NULL;
- smpi_received[i] = NULL;
- smpi_last_received[i] = NULL;
- }
- smpi_timer = xbt_os_timer_new();
- smpi_reference = DEFAULT_POWER;
- smpi_benchmarking = 0;
-
- // tell send/recv nodes to begin
- for(i = 0; i < size; i++) {
- mtask = MSG_task_create("READY", 0, 0, NULL);
- MSG_task_put(mtask, hosts[i], SEND_SYNC_PORT);
- mtask = (m_task_t)0;
- MSG_task_get_from_host(&mtask, SEND_SYNC_PORT, hosts[i]);
- MSG_task_destroy(mtask);
- mtask = MSG_task_create("READY", 0, 0, NULL);
- MSG_task_put(mtask, hosts[i], RECV_SYNC_PORT);
- mtask = (m_task_t)0;
- MSG_task_get_from_host(&mtask, RECV_SYNC_PORT, hosts[i]);
- MSG_task_destroy(mtask);
- }
-
- // now everyone else
- for(i = 1; i < size; i++) {
- mtask = MSG_task_create("READY", 0, 0, NULL);
- MSG_task_put(mtask, hosts[i], MPI_PORT);
- }
-
- } else {
- // everyone needs to wait for node 0 to finish
- mtask = (m_task_t)0;
- MSG_task_get(&mtask, MPI_PORT);
- MSG_task_destroy(mtask);
- smpi_mpi_comm_world.processes[rank] = SIMIX_process_self();
- }
-
- // now that mpi_comm_world_processes is set, it's safe to set a barrier
- smpi_barrier(&smpi_mpi_comm_world);
-}
+XBT_LOG_NEW_DEFAULT_CATEGORY(smpi, "SMPI");
-void smpi_mpi_finalize() {
- int i;
- smpi_running_hosts--;
- if (0 <= smpi_running_hosts) {
- 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]);
- }
- }
- } else {
- xbt_free(smpi_mpi_comm_world.processes);
- xbt_free(smpi_pending_send_requests);
- xbt_free(smpi_last_pending_send_requests);
- xbt_free(smpi_pending_recv_requests);
- xbt_free(smpi_last_pending_recv_requests);
- xbt_free(smpi_received);
- xbt_free(smpi_last_received);
- xbt_free(smpi_sender_processes);
- xbt_free(smpi_receiver_processes);
- xbt_os_timer_free(smpi_timer);
- }
+int inline smpi_mpi_comm_size(smpi_mpi_communicator_t *comm)
+{
+ return comm->size;
}
-void smpi_complete(smpi_mpi_request_t *request) {
- smpi_waitlist_node_t *current, *next;
- request->completed = 1;
- request->next = NULL;
- current = request->waitlist;
- while(NULL != current) {
- if(SIMIX_process_is_suspended(current->process)) {
- SIMIX_process_resume(current->process);
- }
- next = current->next;
- xbt_free(current);
- current = next;
- }
- request->waitlist = NULL;
-}
+// FIXME: smarter algorithm?
+int smpi_mpi_comm_rank(smpi_mpi_communicator_t *comm, smx_host_t host)
+{
+ int i;
-int smpi_host_rank_self() {
- return smpi_comm_rank(&smpi_mpi_comm_world, SIMIX_host_self());
-}
+ for(i = comm->size - 1; i > 0 && host != comm->hosts[i]; i--);
-void smpi_isend(smpi_mpi_request_t *sendreq) {
- int rank = smpi_host_rank_self();
- if (NULL == smpi_last_pending_send_requests[rank]) {
- smpi_pending_send_requests[rank] = sendreq;
- } else {
- smpi_last_pending_send_requests[rank]->next = sendreq;
- }
- smpi_last_pending_send_requests[rank] = sendreq;
- if (SIMIX_process_is_suspended(smpi_sender_processes[rank])) {
- SIMIX_process_resume(smpi_sender_processes[rank]);
- }
+ return i;
}
-void smpi_match_requests(int rank) {
- smpi_mpi_request_t *frequest, *prequest, *crequest;
- smpi_received_t *freceived, *preceived, *creceived;
- size_t dsize;
- short int match;
- frequest = smpi_pending_recv_requests[rank];
- prequest = NULL;
- crequest = frequest;
- while(NULL != crequest) {
- freceived = smpi_received[rank];
- preceived = NULL;
- creceived = freceived;
- match = 0;
- while(NULL != creceived && !match) {
- if(crequest->comm->id == creceived->commid &&
- (MPI_ANY_SOURCE == crequest->src || crequest->src == creceived->src) &&
- crequest->tag == creceived->tag) {
-
- // we have a match!
- match = 1;
-
- // pull the request from the queue
- if(NULL == prequest) {
- frequest = crequest->next;
- smpi_pending_recv_requests[rank] = frequest;
- } else {
- prequest->next = crequest->next;
- }
- if(crequest == smpi_last_pending_recv_requests[rank]) {
- smpi_last_pending_recv_requests[rank] = prequest;
- }
-
- // pull the received data from the queue
- if(NULL == preceived) {
- freceived = creceived->next;
- smpi_received[rank] = freceived;
- } else {
- preceived->next = creceived->next;
- }
- if(creceived == smpi_last_received[rank]) {
- smpi_last_received[rank] = preceived;
- }
-
- // for when request->src is any source
- crequest->src = creceived->src;
-
- // calculate data size
- dsize = crequest->count * crequest->datatype->size;
-
- // copy data to buffer
- memcpy(crequest->buf, creceived->data, dsize);
-
- // fwd through
- crequest->fwdthrough = creceived->fwdthrough;
-
- // get rid of received data node, no longer needed
- xbt_free(creceived->data);
- xbt_free(creceived);
-
- if (crequest->fwdthrough == rank) {
- smpi_complete(crequest);
- } else {
- crequest->src = rank;
- crequest->dst = (rank + 1) % crequest->comm->size;
- smpi_isend(crequest);
- }
-
- } else {
- preceived = creceived;
- creceived = creceived->next;
- }
- }
- prequest = crequest;
- crequest = crequest->next;
- }
+int inline smpi_mpi_comm_rank_self(smpi_mpi_communicator_t *comm)
+{
+ return smpi_mpi_comm_rank(comm, SIMIX_host_self());
}
-void smpi_bench_begin() {
- xbt_assert0(!smpi_benchmarking, "Already benchmarking");
- smpi_benchmarking = 1;
- xbt_os_timer_start(smpi_timer);
- return;
+int inline smpi_mpi_comm_world_rank_self()
+{
+ return smpi_mpi_comm_rank(&smpi_mpi_comm_world, SIMIX_host_self());
}
-void smpi_bench_end() {
- m_task_t ctask = NULL;
- double duration;
- xbt_assert0(smpi_benchmarking, "Not benchmarking yet");
- smpi_benchmarking = 0;
- xbt_os_timer_stop(smpi_timer);
- duration = xbt_os_timer_elapsed(smpi_timer);
- ctask = MSG_task_create("computation", duration * smpi_reference, 0 , NULL);
- MSG_task_execute(ctask);
- MSG_task_destroy(ctask);
- return;
+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_received_message_t *scratch;
+ int drank;
+ smx_process_t waitproc;
+
+ self = SIMIX_process_self();
+ shost = SIMIX_host_self();
+ rank = smpi_mpi_comm_rank(&smpi_mpi_comm_world, shost);
+
+ // make sure root is done before own initialization
+ SIMIX_mutex_lock(init_mutex);
+ if (!smpi_root_ready) {
+ SIMIX_cond_wait(init_cond, init_mutex);
+ }
+ SIMIX_mutex_unlock(init_mutex);
+
+ 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
+ SIMIX_mutex_lock(init_mutex);
+ smpi_ready_count++;
+ if (smpi_ready_count < 3 * size) {
+ SIMIX_cond_wait(init_cond, init_mutex);
+ } else {
+ SIMIX_cond_broadcast(init_cond);
+ }
+ SIMIX_mutex_unlock(init_mutex);
+
+ SIMIX_mutex_lock(smpi_running_hosts_mutex);
+ running_hosts = smpi_running_hosts;
+ SIMIX_mutex_unlock(smpi_running_hosts_mutex);
+
+ while (0 < running_hosts) {
+
+ 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);
+
+ dhost = request->comm->hosts[request->dst];
+
+ // FIXME: not at all sure I can assume magic just happens here....
+ communicate_action = SIMIX_action_communicate(shost, dhost,
+ "communication", request->datatype->size * request->count * 1.0, -1.0);
+
+ SIMIX_register_condition_to_action(communicate_action, request->cond);
+ SIMIX_register_action_to_condition(communicate_action, request->cond);
+
+ 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;
+ drank = smpi_mpi_comm_rank(&smpi_mpi_comm_world, dhost);
+ SIMIX_mutex_lock(smpi_received_messages_mutex[drank]);
+ xbt_fifo_push(smpi_received_messages[drank], scratch);
+ SIMIX_mutex_unlock(smpi_received_messages_mutex[drank]);
+
+ request->completed = 1;
+
+ // 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_lock(smpi_running_hosts_mutex);
+ running_hosts = smpi_running_hosts;
+ SIMIX_mutex_unlock(smpi_running_hosts_mutex);
+ }
+
+ return 0;
}
-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 (NULL == buf && 0 < count) {
- retval = MPI_ERR_INTERN;
- } else if (0 > count) {
- retval = MPI_ERR_COUNT;
- } 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_malloc(sizeof(smpi_mpi_request_t));
- (*request)->buf = buf;
- (*request)->count = count;
- (*request)->datatype = datatype;
- (*request)->src = src;
- (*request)->dst = dst;
- (*request)->tag = tag;
- (*request)->comm = comm;
- (*request)->completed = 0;
- (*request)->fwdthrough = dst;
- (*request)->waitlist = NULL;
- (*request)->next = NULL;
- }
- return retval;
+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;
+ xbt_fifo_item_t request_item, message_item;
+ smpi_mpi_request_t *request;
+ smpi_received_message_t *message;
+ smx_process_t waitproc;
+
+ self = SIMIX_process_self();
+ rank = smpi_mpi_comm_world_rank_self();
+
+ // make sure root is done before own initialization
+ SIMIX_mutex_lock(init_mutex);
+ if (!smpi_root_ready) {
+ SIMIX_cond_wait(init_cond, init_mutex);
+ }
+ SIMIX_mutex_unlock(init_mutex);
+
+ 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
+ SIMIX_mutex_lock(init_mutex);
+ smpi_ready_count++;
+ if (smpi_ready_count < 3 * size) {
+ SIMIX_cond_wait(init_cond, init_mutex);
+ } else {
+ SIMIX_cond_broadcast(init_cond);
+ }
+ SIMIX_mutex_unlock(init_mutex);
+
+ SIMIX_mutex_lock(smpi_running_hosts_mutex);
+ running_hosts = smpi_running_hosts;
+ SIMIX_mutex_unlock(smpi_running_hosts_mutex);
+
+ while (0 < running_hosts) {
+
+ // 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->mutex);
+ memcpy(request->buf, message->buf, request->count * request->datatype->size);
+ request->src = message->src;
+ request->completed = 1;
+
+ while (waitproc = xbt_fifo_shift(request->waitlist)) {
+ if (SIMIX_process_is_suspended(waitproc)) {
+ SIMIX_process_resume(waitproc);
+ }
+ }
+ SIMIX_mutex_unlock(request->mutex);
+
+ xbt_mallocator_release(smpi_message_mallocator, message);
+ }
+
+ SIMIX_mutex_lock(smpi_running_hosts_mutex);
+ running_hosts = smpi_running_hosts;
+ SIMIX_mutex_unlock(smpi_running_hosts_mutex);
+ }
+
+ return 0;
}
-void smpi_barrier(smpi_mpi_communicator_t *comm) {
- int i;
- comm->barrier++;
- if(comm->barrier < comm->size) {
- SIMIX_process_suspend(SIMIX_process_self());
- } else {
- comm->barrier = 0;
- for(i = 0; i < comm->size; i++) {
- if (SIMIX_process_is_suspended(comm->processes[i])) {
- SIMIX_process_resume(comm->processes[i]);
- }
- }
- }
+int smpi_run_simulation(int argc, char **argv)
+{
+ 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);
+
+ 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_create_environment(argv[1]);
+ SIMIX_launch_application(argv[2]);
+
+ /* Prepare to display some more info when dying on Ctrl-C pressing */
+ //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());
+ SIMIX_clean();
+ return 0;
}
-int smpi_comm_rank(smpi_mpi_communicator_t *comm, smx_host_t host) {
- int i;
- for(i = 0; i < comm->size && host != comm->hosts[i]; i++);
- if (i >= comm->size) i = -1;
- return i;
+void smpi_mpi_land_func(void *x, void *y, void *z)
+{
+ *(int *)z = *(int *)x && *(int *)y;
}
-void smpi_irecv(smpi_mpi_request_t *recvreq) {
- int rank = smpi_host_rank_self();
- if (NULL == smpi_pending_recv_requests[rank]) {
- smpi_pending_recv_requests[rank] = recvreq;
- } else if (NULL != smpi_last_pending_recv_requests[rank]) {
- smpi_last_pending_recv_requests[rank]->next = recvreq;
- } else { // can't happen!
- fprintf(stderr, "smpi_pending_recv_requests not null while smpi_last_pending_recv_requests null!\n");
- }
- smpi_last_pending_recv_requests[rank] = recvreq;
- smpi_match_requests(rank);
- if (SIMIX_process_is_suspended(smpi_receiver_processes[rank])) {
- SIMIX_process_resume(smpi_receiver_processes[rank]);
- }
+void smpi_mpi_sum_func(void *x, void *y, void *z)
+{
+ *(int *)z = *(int *)x + *(int *)y;
}
-void smpi_wait(smpi_mpi_request_t *request, smpi_mpi_status_t *status) {
- smpi_waitlist_node_t *waitnode, *current;
- if (NULL != request) {
- if (!request->completed) {
- waitnode = xbt_malloc(sizeof(smpi_waitlist_node_t));
- waitnode->process = SIMIX_process_self();
- waitnode->next = NULL;
- if (NULL == request->waitlist) {
- request->waitlist = waitnode;
- } else {
- for(current = request->waitlist; NULL != current->next; current = current->next);
- current->next = waitnode;
- }
- SIMIX_process_suspend(waitnode->process);
- }
- if (NULL != status && MPI_STATUS_IGNORE != status) {
- status->MPI_SOURCE = request->src;
- }
- }
+void *smpi_new_request()
+{
+ return xbt_new(smpi_mpi_request_t, 1);
}
-void smpi_wait_all(int count, smpi_mpi_request_t **requests, smpi_mpi_status_t *statuses) {
- int i;
- for (i = 0; i < count; i++) {
- smpi_wait(requests[i], &statuses[i]);
- }
+void *smpi_new_message()
+{
+ return xbt_new(smpi_received_message_t, 1);
}
-void smpi_wait_all_nostatus(int count, smpi_mpi_request_t **requests) {
- int i;
- for (i = 0; i < count; i++) {
- smpi_wait(requests[i], MPI_STATUS_IGNORE);
- }
-}
+void smpi_mpi_init()
+{
+ int i;
+ int size;
+ smx_process_t process;
+ smx_host_t *hosts;
+ smx_host_t host;
+ double duration;
+
+ // initialize some local variables
+ host = SIMIX_host_self();
+ hosts = SIMIX_host_get_table();
+ size = SIMIX_host_get_number();
+
+ // node 0 sets the globals
+ if (host == hosts[0]) {
+
+ // processes
+ smpi_sender_processes = xbt_new(smx_process_t, size);
+ smpi_receiver_processes = xbt_new(smx_process_t, size);
+
+ // running hosts
+ smpi_running_hosts_mutex = SIMIX_mutex_init();
+ smpi_running_hosts = size;
+
+ // global communicator
+ smpi_mpi_comm_world.size = size;
+ smpi_mpi_comm_world.barrier = 0;
+ smpi_mpi_comm_world.barrier_mutex = SIMIX_mutex_init();
+ smpi_mpi_comm_world.barrier_cond = SIMIX_cond_init();
+ smpi_mpi_comm_world.hosts = hosts;
+ smpi_mpi_comm_world.processes = xbt_new(smx_process_t, size);
+ smpi_mpi_comm_world.processes[0] = SIMIX_process_self();
+
+ // mpi datatypes
+ smpi_mpi_byte.size = (size_t)1;
+ smpi_mpi_int.size = sizeof(int);
+ smpi_mpi_double.size = sizeof(double);
+
+ // mpi operations
+ smpi_mpi_land.func = &smpi_mpi_land_func;
+ 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_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_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();
+ smpi_reference_speed = SMPI_DEFAULT_SPEED;
+ smpi_benchmarking = 0;
+ smpi_benchmarking_mutex = SIMIX_mutex_init();
+
+ // signal all nodes to perform initialization
+ SIMIX_mutex_lock(init_mutex);
+ smpi_root_ready = 1;
+ SIMIX_cond_broadcast(init_cond);
+ SIMIX_mutex_unlock(init_mutex);
+
+ } else {
+
+ // make sure root is done before own initialization
+ SIMIX_mutex_lock(init_mutex);
+ if (!smpi_root_ready) {
+ SIMIX_cond_wait(init_cond, init_mutex);
+ }
+ SIMIX_mutex_unlock(init_mutex);
+
+ smpi_mpi_comm_world.processes[smpi_mpi_comm_rank_self(&smpi_mpi_comm_world)] = SIMIX_process_self();
+
+ }
+
+ // wait for all nodes to signal initializatin complete
+ SIMIX_mutex_lock(init_mutex);
+ smpi_ready_count++;
+ if (smpi_ready_count < 3 * size) {
+ SIMIX_cond_wait(init_cond, init_mutex);
+ } else {
+ SIMIX_cond_broadcast(init_cond);
+ }
+ SIMIX_mutex_unlock(init_mutex);
-int smpi_sender(int argc, char *argv[]) {
- smx_process_t process;
- char taskname[50];
- size_t dsize;
- void *data;
- smx_host_t dhost;
- m_task_t mtask;
- int rank, fc, ft;
- smpi_mpi_request_t *sendreq;
-
- process = SIMIX_process_self();
-
- // wait for init
- mtask = (m_task_t)0;
- MSG_task_get(&mtask, SEND_SYNC_PORT);
-
- rank = smpi_host_rank_self();
-
- smpi_sender_processes[rank] = process;
-
- // ready!
- MSG_task_put(mtask, MSG_task_get_source(mtask), SEND_SYNC_PORT);
-
- while (0 < smpi_running_hosts) {
- sendreq = smpi_pending_send_requests[rank];
- if (NULL != sendreq) {
-
- // pull from queue if not a fwd or no more to fwd
- if (sendreq->dst == sendreq->fwdthrough) {
- smpi_pending_send_requests[rank] = sendreq->next;
- if(sendreq == smpi_last_pending_send_requests[rank]) {
- smpi_last_pending_send_requests[rank] = NULL;
- }
- ft = sendreq->dst;
- } else {
- fc = ((sendreq->fwdthrough - sendreq->dst + sendreq->comm->size) % sendreq->comm->size) / 2;
- ft = (sendreq->dst + fc) % sendreq->comm->size;
- //printf("node %d sending broadcast to node %d through node %d\n", rank, sendreq->dst, ft);
- }
-
- // create task to send
- sprintf(taskname, "comm:%d,src:%d,dst:%d,tag:%d,ft:%d", sendreq->comm->id, sendreq->src, sendreq->dst, sendreq->tag, ft);
- dsize = sendreq->count * sendreq->datatype->size;
- data = xbt_malloc(dsize);
- memcpy(data, sendreq->buf, dsize);
- mtask = MSG_task_create(taskname, 0, dsize, data);
-
- // figure out which host to send it to
- dhost = sendreq->comm->hosts[sendreq->dst];
-
- // send task
- #ifdef DEBUG
- printf("host %s attempting to send to host %s\n", SIMIX_host_get_name(SIMIX_host_self()), SIMIX_host_get_name(dhost));
- #endif
- MSG_task_put(mtask, dhost, MPI_PORT);
-
- if (sendreq->dst == sendreq->fwdthrough) {
- smpi_complete(sendreq);
- } else {
- sendreq->dst = (sendreq->dst + fc + 1) % sendreq->comm->size;
- }
-
- } else {
- SIMIX_process_suspend(process);
- }
- }
- return 0;
}
-int smpi_receiver(int argc, char **argv) {
- smx_process_t process;
- m_task_t mtask;
- smpi_received_t *received;
- int rank;
- smpi_mpi_request_t *recvreq;
+void smpi_mpi_finalize()
+{
+ int i;
- process = SIMIX_process_self();
+ SIMIX_mutex_lock(smpi_running_hosts_mutex);
+ i = --smpi_running_hosts;
+ SIMIX_mutex_unlock(smpi_running_hosts_mutex);
- // wait for init
- mtask = (m_task_t)0;
- MSG_task_get(&mtask, RECV_SYNC_PORT);
+ if (0 >= i) {
- rank = smpi_host_rank_self();
+ SIMIX_mutex_destroy(smpi_running_hosts_mutex);
- // potential race condition...
- smpi_receiver_processes[rank] = process;
+ 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]);
+ }
- // ready!
- MSG_task_put(mtask, MSG_task_get_source(mtask), RECV_SYNC_PORT);
+ 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);
- while (0 < smpi_running_hosts) {
- recvreq = smpi_pending_recv_requests[rank];
- if (NULL != recvreq) {
- mtask = (m_task_t)0;
+ SIMIX_mutex_destroy(smpi_mpi_comm_world.barrier_mutex);
+ SIMIX_cond_destroy(smpi_mpi_comm_world.barrier_cond);
+ xbt_free(smpi_mpi_comm_world.processes);
- #ifdef DEBUG
- printf("host %s waiting to receive from anyone, but first in queue is (%d,%d,%d).\n",
- SIMIX_host_get_name(SIMIX_host_self()), recvreq->src, recvreq->dst, recvreq->tag);
- #endif
- MSG_task_get(&mtask, MPI_PORT);
+ xbt_os_timer_free(smpi_timer);
+ }
- received = xbt_malloc(sizeof(smpi_received_t));
+}
- sscanf(MSG_task_get_name(mtask), "comm:%d,src:%d,dst:%d,tag:%d,ft:%d",
- &received->commid, &received->src, &received->dst, &received->tag, &received->fwdthrough);
- received->data = MSG_task_get_data(mtask);
- received->next = NULL;
+void smpi_bench_begin()
+{
+ xbt_assert0(!smpi_benchmarking, "Already benchmarking");
+ smpi_benchmarking = 1;
+ xbt_os_timer_start(smpi_timer);
+ return;
+}
- if (NULL == smpi_last_received[rank]) {
- smpi_received[rank] = received;
- } else {
- smpi_last_received[rank]->next = received;
- }
- smpi_last_received[rank] = received;
+void smpi_bench_end()
+{
+ double duration;
+ smx_host_t host;
+ smx_action_t compute_action;
+ smx_mutex_t mutex;
+ smx_cond_t cond;
+
+ xbt_assert0(smpi_benchmarking, "Not benchmarking yet");
+ smpi_benchmarking = 0;
+ xbt_os_timer_stop(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_cond_wait(cond, mutex);
+ SIMIX_mutex_unlock(mutex);
+ SIMIX_mutex_destroy(mutex);
+ SIMIX_cond_destroy(cond);
+ // FIXME: check for success/failure?
+ return;
+}
- MSG_task_destroy(mtask);
+void smpi_barrier(smpi_mpi_communicator_t *comm) {
+ int i;
+ SIMIX_mutex_lock(comm->barrier_mutex);
+ comm->barrier++;
+ if(i < comm->size) {
+ SIMIX_cond_wait(comm->barrier_cond, comm->barrier_mutex);
+ } else {
+ comm->barrier = 0;
+ SIMIX_cond_broadcast(comm->barrier_cond);
+ }
+ SIMIX_mutex_unlock(comm->barrier_mutex);
+}
+
+int smpi_comm_rank(smpi_mpi_communicator_t *comm, smx_host_t host)
+{
+ 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_request_mallocator);
+ (*request)->comm = comm;
+ (*request)->src = src;
+ (*request)->dst = dst;
+ (*request)->tag = tag;
+ (*request)->buf = buf;
+ (*request)->count = count;
+ (*request)->datatype = datatype;
+ (*request)->completed = 0;
+ (*request)->mutex = SIMIX_mutex_init();
+ (*request)->cond = SIMIX_cond_init();
+ (*request)->waitlist = NULL;
+ }
+ return retval;
+}
- smpi_match_requests(rank);
+int smpi_isend(smpi_mpi_request_t *request)
+{
+ int rank = smpi_mpi_comm_rank_self(&smpi_mpi_comm_world);
+
+ 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]);
+ }
+}
+
+int smpi_irecv(smpi_mpi_request_t *request)
+{
+ int rank = smpi_mpi_comm_rank_self(&smpi_mpi_comm_world);
+
+ 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]);
+ }
+}
- } else {
- SIMIX_process_suspend(process);
- }
- }
- return 0;
+void smpi_wait(smpi_mpi_request_t *request, smpi_mpi_status_t *status)
+{
+ smx_process_t self;
+ int suspend = 0;
+ self = SIMIX_process_self();
+
+ if (NULL != request) {
+ SIMIX_mutex_lock(request->mutex);
+ if (!request->completed) {
+ xbt_fifo_push(request->waitlist, self);
+ suspend = 1;
+ }
+ SIMIX_mutex_unlock(request->mutex);
+ if (suspend) {
+ SIMIX_suspend(self);
+ }
+ if (NULL != status && MPI_STATUS_IGNORE != status) {
+ SIMIX_mutex_lock(request->mutex);
+ status->MPI_SOURCE = request->src;
+ SIMIX_mutex_unlock(request->mutex);
+ }
+ }
}
// 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);
- }
- smpi_bench_begin();
- return retval;
+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);
+ }
+ smpi_bench_begin();
+ return retval;
}
-unsigned int smpi_sleep(unsigned int seconds) {
- m_task_t task = NULL;
- smpi_bench_end();
- task = MSG_task_create("sleep", seconds * DEFAULT_POWER, 0, NULL);
- MSG_task_execute(task);
- MSG_task_destroy(task);
- smpi_bench_begin();
- return 0;
+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_condition_to_action(sleep_action, cond);
+ SIMIX_register_action_to_condition(sleep_action, cond);
+ SIMIX_cond_wait(cond, mutex);
+ SIMIX_mutex_unlock(mutex);
+ SIMIX_mutex_destroy(mutex);
+ SIMIX_cond_destroy(cond);
+ // FIXME: check for success/failure?
+ smpi_bench_begin();
+ return 0;
}
-void smpi_exit(int status) {
- smpi_bench_end();
- smpi_running_hosts--;
- SIMIX_process_kill(SIMIX_process_self());
- return;
+void smpi_exit(int status)
+{
+ smpi_bench_end();
+ SIMIX_mutex_lock(smpi_running_hosts_mutex);
+ smpi_running_hosts--;
+ SIMIX_mutex_unlock(smpi_running_hosts_mutex);
+ SIMIX_process_kill(SIMIX_process_self());
+ return;
}
