From: Augustin Degomme Date: Wed, 12 Dec 2012 18:33:40 +0000 (+0100) Subject: change the way we handle MPI_Request termination X-Git-Tag: v3_9_rc1~86^2~118 X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/commitdiff_plain/7b5d97a0198ce4ecbe5654a7cfe978bd1880f936 change the way we handle MPI_Request termination to avoid using smx_action too much in smpi. Removes a few more leaks in collectives Handling of waitany was changed, it should return sooner and avoid a few unnecessary simcalls --- diff --git a/src/smpi/private.h b/src/smpi/private.h index 3831ae2b68..c23edc6ad3 100644 --- a/src/smpi/private.h +++ b/src/smpi/private.h @@ -69,6 +69,8 @@ typedef struct s_smpi_mpi_request { smx_action_t action; unsigned flags; int detached; + MPI_Request detached_sender; + int refcount; #ifdef HAVE_TRACING int send; int recv; diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index e548adb854..594e4b74ca 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -30,7 +30,9 @@ static int match_recv(void* a, void* b, smx_action_t ignored) { 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; - else ref->truncated = 0; + 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; } @@ -48,7 +50,10 @@ static int match_send(void* a, void* b,smx_action_t ignored) { 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; - else ref->truncated = 0; + 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; } @@ -88,6 +93,13 @@ static MPI_Request build_request(void *buf, int count, 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; @@ -148,7 +160,7 @@ MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype, MPI_Request request = build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, comm, PERSISTENT | SEND); - + request->refcount++; return request; } @@ -158,7 +170,7 @@ MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype, MPI_Request request = build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag, comm, PERSISTENT | RECV); - + request->refcount++; return request; } @@ -178,7 +190,6 @@ void smpi_mpi_start(MPI_Request request) 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); - if (request->action)request->action->comm.refcount++; } else { int receiver = smpi_group_index(smpi_comm_group(request->comm), request->dst); @@ -195,9 +206,10 @@ void smpi_mpi_start(MPI_Request request) } if (request->size < 64*1024 ) { //(FIXME: this limit should be configurable) void *oldbuf = NULL; + request->detached = 1; + request->refcount++; if(request->old_type->has_subtype == 0){ oldbuf = request->buf; - request->detached = 1; if (oldbuf){ request->buf = malloc(request->size); memcpy(request->buf,oldbuf,request->size); @@ -217,7 +229,6 @@ void smpi_mpi_start(MPI_Request request) request, // detach if msg size < eager/rdv switch limit request->detached); - if (request->action)request->action->comm.refcount++; #ifdef HAVE_TRACING /* FIXME: detached sends are not traceable (request->action == NULL) */ @@ -240,8 +251,18 @@ void smpi_mpi_startall(int count, MPI_Request * requests) void smpi_mpi_request_free(MPI_Request * request) { - xbt_free(*request); - *request = MPI_REQUEST_NULL; + + 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"); + } } MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, @@ -258,7 +279,8 @@ MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = - smpi_isend_init(buf, count, datatype, dst, tag, comm); + build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, + comm, NON_PERSISTENT | SEND); smpi_mpi_start(request); return request; @@ -277,7 +299,8 @@ MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) { MPI_Request request = - smpi_irecv_init(buf, count, datatype, src, tag, comm); + build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag, + comm, NON_PERSISTENT | RECV); smpi_mpi_start(request); return request; @@ -296,11 +319,8 @@ void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, 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 | RECV_DELETE); - - smpi_mpi_start(request); + MPI_Request request; + request = smpi_mpi_isend(buf, count, datatype, dst, tag, comm); smpi_mpi_wait(&request, MPI_STATUS_IGNORE); } @@ -333,74 +353,41 @@ int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype) static void finish_wait(MPI_Request * request, MPI_Status * status) { MPI_Request req = *request; - - 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; - if(req->truncated) - status->MPI_ERROR = MPI_ERR_TRUNCATE; - else status->MPI_ERROR = 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; - } - req = *request; - - 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->flags & NON_PERSISTENT) { - if(req->flags & RECV && - req->action && - (req->action->state == SIMIX_DONE)) - { - MPI_Request sender_request = (MPI_Request)SIMIX_comm_get_src_data(req->action); - if((sender_request!=MPI_REQUEST_NULL) && - ( sender_request->detached ) && - ( sender_request->flags & RECV_DELETE)) - { - //we are in a receiver's wait from a detached send - //we have to clean the sender's side request here.... but only if done by a send, not an isend - //the request lives senderside for an isend. As detached is currently for send + isend, we use RECV_DELETE to separate them - //FIXME : see if just removing detached status for isend is also good - smpi_mpi_request_free(&sender_request); - } + 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; + if(req->truncated) + status->MPI_ERROR = MPI_ERR_TRUNCATE; + else status->MPI_ERROR = 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(req->action){ - //if we want to free our request, we have to invalidate it at the other end of the comm - if(req->flags & SEND){ - req->action->comm.src_data=MPI_REQUEST_NULL; - }else{ - req->action->comm.dst_data=MPI_REQUEST_NULL; + 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); } - - smx_action_t temp=req->action; - if(req->action->comm.refcount == 1)req->action = NULL; - SIMIX_comm_destroy(temp); + 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 { - if(req->action)SIMIX_comm_destroy(req->action); req->action = NULL; } } @@ -414,6 +401,7 @@ int smpi_mpi_test(MPI_Request * request, MPI_Status * status) { else flag = simcall_comm_test((*request)->action); if(flag) { + (*request)->refcount++; finish_wait(request, status); }else{ smpi_empty_status(status); @@ -538,8 +526,8 @@ 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); } + finish_wait(request, status); // FIXME for a detached send, finish_wait is not called: } @@ -557,13 +545,22 @@ int smpi_mpi_waitany(int count, MPI_Request requests[], comms = xbt_dynar_new(sizeof(smx_action_t), NULL); map = xbt_new(int, count); size = 0; - XBT_DEBUG("Wait for one of"); + XBT_DEBUG("Wait for one of %d", count); for(i = 0; i < count; i++) { - if((requests[i] != MPI_REQUEST_NULL) && (requests[i]->action != NULL)) { - print_request("Waiting any ", requests[i]); - xbt_dynar_push(comms, &requests[i]->action); - map[size] = i; - size++; + 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) { @@ -604,7 +601,6 @@ int smpi_mpi_waitall(int count, MPI_Request requests[], } } } - for(c = 0; c < count; c++) { if(MC_is_active()) { smpi_mpi_wait(&requests[c], pstat); @@ -621,6 +617,7 @@ int smpi_mpi_waitall(int count, MPI_Request requests[], } } } + return retvalue; } diff --git a/src/smpi/smpi_pmpi.c b/src/smpi/smpi_pmpi.c index 9fde3a03f9..1e4dc070af 100644 --- a/src/smpi/smpi_pmpi.c +++ b/src/smpi/smpi_pmpi.c @@ -897,9 +897,10 @@ int PMPI_Request_free(MPI_Request * request) int retval; smpi_bench_end(); - if (request == MPI_REQUEST_NULL) { + if (*request == MPI_REQUEST_NULL) { retval = MPI_ERR_ARG; } else { + if((*request)->flags & PERSISTENT)(*request)->refcount--; smpi_mpi_request_free(request); retval = MPI_SUCCESS; } diff --git a/teshsuite/smpi/mpich-test/pt2pt/sndrcv.c b/teshsuite/smpi/mpich-test/pt2pt/sndrcv.c index 9323ccc1df..10b418a1f7 100644 --- a/teshsuite/smpi/mpich-test/pt2pt/sndrcv.c +++ b/teshsuite/smpi/mpich-test/pt2pt/sndrcv.c @@ -95,20 +95,20 @@ for (i=0; i