X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/aa64b0469b0967780beb549867c2907c14525443..6b9a59680b27bbe789336585bd00be1fae84d9a7:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index b4e8aa6841..a016c1dafa 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -5,10 +5,11 @@ * under the terms of the license (GNU LGPL) which comes with this package. */ #include "private.h" -#include "xbt/time.h" +#include "xbt/virtu.h" #include "mc/mc.h" #include "xbt/replay.h" #include +#include "simix/smx_private.h" #include "surf/surf.h" @@ -16,23 +17,40 @@ 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; + 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"); - return (ref->src == MPI_ANY_SOURCE || req->src == ref->src) - && (ref->tag == MPI_ANY_TAG || req->tag == ref->tag); + 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; + else ref->truncated = 0; + return 1; + }else return 0; } 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_assert(ref, "Cannot match send against null reference"); - xbt_assert(req, "Cannot match send against null request"); - return (req->src == MPI_ANY_SOURCE || req->src == ref->src) - && (req->tag == MPI_ANY_TAG || req->tag == ref->tag); + 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; + else ref->truncated = 0; + return 1; + } else return 0; } static MPI_Request build_request(void *buf, int count, @@ -41,14 +59,14 @@ static MPI_Request build_request(void *buf, int count, { MPI_Request request; - void *old_buf; + 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-contignous memory + // This part handles the problem of non-contiguous memory old_buf = buf; buf = malloc(count*smpi_datatype_size(datatype)); if (flags & SEND) { @@ -57,7 +75,7 @@ static MPI_Request build_request(void *buf, int count, } request->buf = buf; - // This part handles the problem of non-contignous memory (for the + // This part handles the problem of non-contiguous memory (for the // unserialisation at the reception) request->old_buf = old_buf; request->old_type = datatype; @@ -69,6 +87,7 @@ static MPI_Request build_request(void *buf, int count, request->comm = comm; request->action = NULL; request->flags = flags; + request->detached = 0; #ifdef HAVE_TRACING request->send = 0; request->recv = 0; @@ -76,6 +95,15 @@ static MPI_Request build_request(void *buf, int count, 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->count=0; + } +} + void smpi_action_trace_run(char *path) { char *name; @@ -135,47 +163,57 @@ MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype, void smpi_mpi_start(MPI_Request request) { smx_rdv_t mailbox; - int detached = 0; xbt_assert(!request->action, "Cannot (re)start a non-finished communication"); if(request->flags & RECV) { print_request("New recv", request); - if (request->size < xbt_cfg_get_int(_surf_cfg_set, "smpi/async_small_thres")) + if (request->size < surf_cfg_get_int("smpi/async_small_thres")) mailbox = smpi_process_mailbox_small(); else mailbox = smpi_process_mailbox(); - - // FIXME: SIMIX does not yet support non-contiguous datatypes - request->action = simcall_comm_irecv(mailbox, request->buf, &request->size, &match_recv, request); + // 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; + request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request); + if (request->action)request->action->comm.refcount++; } else { - print_request("New send", request); - if (request->size < xbt_cfg_get_int(_surf_cfg_set, "smpi/async_small_thres")) { // eager mode => detached send (FIXME: this limit should be configurable) - mailbox = smpi_process_remote_mailbox_small( - smpi_group_index(smpi_comm_group(request->comm), request->dst)); + 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 < surf_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( - smpi_group_index(smpi_comm_group(request->comm), request->dst)); + mailbox = smpi_process_remote_mailbox(receiver); } if (request->size < 64*1024 ) { //(FIXME: this limit should be configurable) - void *oldbuf = request->buf; - detached = 1; - request->buf = malloc(request->size); - if (oldbuf) - memcpy(request->buf,oldbuf,request->size); + void *oldbuf = NULL; + 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); + } + } 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; request->action = simcall_comm_isend(mailbox, request->size, -1.0, - request->buf, request->size, + 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 - detached); + request->detached); + if (request->action)request->action->comm.refcount++; #ifdef HAVE_TRACING /* FIXME: detached sends are not traceable (request->action == NULL) */ @@ -254,10 +292,13 @@ 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; + MPI_Request request = + build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, + comm, NON_PERSISTENT | SEND | RECV_DELETE); - request = smpi_mpi_isend(buf, count, datatype, dst, tag, comm); + smpi_mpi_start(request); smpi_mpi_wait(&request, MPI_STATUS_IGNORE); + } void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, @@ -288,38 +329,74 @@ 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 we have a sender, we should use its data, and not the data from the receive - if((req->action)&& - (req->src==MPI_ANY_SOURCE || req->tag== MPI_ANY_TAG)) - req = (MPI_Request)SIMIX_comm_get_src_data((*request)->action); if(status != MPI_STATUS_IGNORE) { - status->MPI_SOURCE = req->src; - status->MPI_TAG = req->tag; - status->MPI_ERROR = MPI_SUCCESS; + 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->size; + 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->size/smpi_datatype_size(datatype) , datatype->substruct); + subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct); } - //FIXME: I am not sure that if the send is detached we have to free - //the sender buffer thus I do it only for the reciever - if(req->flags & RECV) free(req->buf); + if(req->detached == 0) free(req->buf); } + + 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->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; + } + + smx_action_t temp=req->action; + if(req->action->comm.refcount == 1)req->action = NULL; + SIMIX_comm_destroy(temp); + } + + + smpi_mpi_request_free(request); + + } else { + if(req->action)SIMIX_comm_destroy(req->action); req->action = NULL; } } @@ -327,12 +404,15 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) 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) { finish_wait(request, status); + }else{ + smpi_empty_status(status); } return flag; } @@ -351,24 +431,29 @@ int smpi_mpi_testany(int count, MPI_Request requests[], int *index, map = xbt_new(int, count); size = 0; for(i = 0; i < count; i++) { - if(requests[i]->action) { - xbt_dynar_push(comms, &requests[i]->action); - map[size] = i; - size++; + 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); - // FIXME: MPI_UNDEFINED or does SIMIX have a return code? - if(i != MPI_UNDEFINED) { + // not MPI_UNDEFINED, as this is a simix return code + if(i != -1) { *index = map[i]; - smpi_mpi_wait(&requests[*index], status); + 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; } @@ -376,12 +461,22 @@ int smpi_mpi_testany(int count, MPI_Request requests[], int *index, 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; i0){ - mailbox = smpi_process_mailbox_small(); - request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request); - - } - if (request->action==NULL){ - mailbox = smpi_process_mailbox(); - request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request); - } + if (surf_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=true; + *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->size; + status->count = req->real_size; } } - else *flag=false; + else *flag = 0; + smpi_mpi_request_free(&request); - return request; + return; } void smpi_mpi_wait(MPI_Request * request, MPI_Status * status) @@ -468,57 +564,111 @@ int smpi_mpi_waitany(int count, MPI_Request requests[], if(size > 0) { i = simcall_comm_waitany(comms); - // FIXME: MPI_UNDEFINED or does SIMIX have a return code? - if (i != MPI_UNDEFINED) { + // 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_mpi_waitall(int count, MPI_Request requests[], +int smpi_mpi_waitall(int count, MPI_Request requests[], MPI_Status status[]) { - int index, c; + int index, c; MPI_Status stat; - MPI_Status *pstat = status == MPI_STATUS_IGNORE ? MPI_STATUS_IGNORE : &stat; - + MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat; + int retvalue=MPI_SUCCESS; + //tag invalid requests in the set for(c = 0; c < count; c++) { - if(MC_IS_ENABLED) { - smpi_mpi_wait(&requests[c], pstat); - index = c; - } else { - index = smpi_mpi_waitany(count, requests, pstat); - if(index == MPI_UNDEFINED) { - break; + if(requests[c]==MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ){ + if(status != MPI_STATUSES_IGNORE) + smpi_empty_status(&status[c]); + }else if(requests[c]->src == MPI_PROC_NULL ){ + if(status != MPI_STATUSES_IGNORE) { + smpi_empty_status(&status[c]); + status[c].MPI_SOURCE=MPI_PROC_NULL; } } - if(status != MPI_STATUS_IGNORE) { - memcpy(&status[index], pstat, sizeof(*pstat)); + } + + 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) { + memcpy(&status[index], pstat, sizeof(*pstat)); + if(status[index].MPI_ERROR==MPI_ERR_TRUNCATE)retvalue=MPI_ERR_IN_STATUS; + + } } } + return retvalue; } int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[]) { 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++) { - if(smpi_mpi_testany(incount, requests, &index, status)) { + 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) { + memcpy(&status[index], pstat, sizeof(*pstat)); + } + }else{ + return MPI_UNDEFINED; } } return count; } +int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, + MPI_Status status[]) +{ + 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) { + memcpy(&status[i], pstat, sizeof(*pstat)); + } + } + }else{ + count_dead++; + } + } + if(count_dead==incount)return MPI_UNDEFINED; + else return count; +} + void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm) { @@ -806,7 +956,7 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, smpi_mpi_startall(size - 1, requests); for(src = 0; src < size - 1; src++) { index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE); - XBT_VERB("finished waiting any request with index %d", index); + XBT_DEBUG("finished waiting any request with index %d", index); if(index == MPI_UNDEFINED) { break; }