X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/281f8c29e177852dcb1591fc31e363e1176857c8..7749bb00b0a00619f255f091ec53f2b304be06d9:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index 952db704fa..38011e60e7 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -1,4 +1,4 @@ -/* Copyright (c) 2007-2013. The SimGrid Team. +/* Copyright (c) 2007-2014. The SimGrid Team. * All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it @@ -197,11 +197,11 @@ static MPI_Request build_request(void *buf, int count, s_smpi_subtype_t *subtype = datatype->substruct; - if(datatype->has_subtype == 1){ + if(((flags & RECV) && (flags & ACCUMULATE)) || (datatype->has_subtype == 1)){ // This part handles the problem of non-contiguous memory old_buf = buf; buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype)); - if (flags & SEND) { + if ((datatype->has_subtype == 1) && (flags & SEND)) { subtype->serialize(old_buf, buf, count, datatype->substruct); } } @@ -226,8 +226,11 @@ static MPI_Request build_request(void *buf, int count, request->truncated = 0; request->real_size = 0; request->real_tag = 0; - - request->refcount=1; + if(flags & PERSISTENT) + request->refcount = 1; + else + request->refcount = 0; + request->op = MPI_REPLACE; #ifdef HAVE_TRACING request->send = 0; request->recv = 0; @@ -257,8 +260,8 @@ void smpi_action_trace_run(char *path) action_fp=NULL; if (path) { action_fp = fopen(path, "r"); - xbt_assert(action_fp != NULL, "Cannot open %s: %s", path, - strerror(errno)); + if (action_fp == NULL) + xbt_die("Cannot open %s: %s", path, strerror(errno)); } if (!xbt_dict_is_empty(action_queues)) { @@ -290,7 +293,6 @@ MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype, MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SEND | PREPARED); - request->refcount++; return request; } @@ -300,7 +302,6 @@ MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype, MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED); - request->refcount++; return request; } @@ -310,7 +311,6 @@ MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype, MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, comm, PERSISTENT | RECV | PREPARED); - request->refcount++; return request; } @@ -320,10 +320,13 @@ void smpi_mpi_start(MPI_Request request) xbt_assert(!request->action, "Cannot (re)start a non-finished communication"); request->flags &= ~PREPARED; + request->flags &= ~FINISHED; + request->refcount++; + if (request->flags & RECV) { print_request("New recv", request); //FIXME: if receive is posted with a large size, but send is smaller, mailboxes may not match ! - if (request->size < sg_cfg_get_int("smpi/async_small_thres")) + if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) mailbox = smpi_process_mailbox_small(); else mailbox = smpi_process_mailbox(); @@ -331,7 +334,10 @@ void smpi_mpi_start(MPI_Request request) request->real_size=request->size; smpi_datatype_use(request->old_type); smpi_comm_use(request->comm); - request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request); + request->action = simcall_comm_irecv(mailbox, request->buf, + &request->real_size, &match_recv, + &smpi_comm_copy_buffer_callback, + request, -1.0); //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; @@ -346,7 +352,7 @@ void smpi_mpi_start(MPI_Request request) int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst); #ifdef HAVE_TRACING - int rank = smpi_process_index(); + int rank = request->src; if (TRACE_smpi_view_internals()) { TRACE_smpi_send(rank, rank, receiver,request->size); } @@ -356,24 +362,33 @@ void smpi_mpi_start(MPI_Request request) /* return;*/ /* }*/ print_request("New send", request); - if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode + if (request->flags & RMA || 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); } + + void* buf = request->buf; if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) { void *oldbuf = NULL; request->detached = 1; + XBT_DEBUG("Send request %p is detached", request); request->refcount++; if(request->old_type->has_subtype == 0){ oldbuf = request->buf; - if (oldbuf && request->size!=0){ - request->buf = xbt_malloc(request->size); - memcpy(request->buf,oldbuf,request->size); + if (!_xbt_replay_is_active() && oldbuf && request->size!=0){ + if((smpi_privatize_global_variables) + && ((char*)request->buf >= start_data_exe) + && ((char*)request->buf < start_data_exe + size_data_exe )){ + XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment "); + switch_data_segment(request->src); + } + buf = xbt_malloc(request->size); + memcpy(buf,oldbuf,request->size); + XBT_DEBUG("buf %p copied into %p",oldbuf,buf); } } - 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 @@ -382,22 +397,22 @@ void smpi_mpi_start(MPI_Request request) smpi_comm_use(request->comm); //if we are giving back the control to the user without waiting for completion, we have to inject timings - double sleeptime =0.0; + 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){ + if(sleeptime != 0.0){ simcall_process_sleep(sleeptime); XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size)); } - request->action = - simcall_comm_isend(mailbox, request->size, -1.0, - request->buf, request->real_size, + simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0, + buf, request->real_size, &match_send, - &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails + &xbt_free_f, // how to free the userdata if a detached send fails + &smpi_comm_copy_buffer_callback, request, // detach if msg size < eager/rdv switch limit request->detached); @@ -435,19 +450,50 @@ void smpi_mpi_request_free(MPI_Request * request) *request = MPI_REQUEST_NULL; }else{ print_request("Decrementing", (*request)); - } }else{ xbt_die("freeing an already free request"); } } + +MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype, + int src, int dst, int tag, MPI_Comm comm, MPI_Op op) +{ + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + if(op==MPI_OP_NULL){ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, src, dst, tag, + comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED); + }else{ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag, + comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE); + request->op = op; + } + return request; +} + +MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype, + int src, int dst, int tag, MPI_Comm comm, MPI_Op op) +{ + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + if(op==MPI_OP_NULL){ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag, + comm, RMA | NON_PERSISTENT | RECV | PREPARED); + }else{ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag, + comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE); + request->op = op; + } + return request; +} + + MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, - comm, NON_PERSISTENT | ISEND | SEND | PREPARED); + comm, PERSISTENT | ISEND | SEND | PREPARED); return request; } @@ -471,14 +517,12 @@ MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype, return request; } - - MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) { MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, - comm, NON_PERSISTENT | RECV | PREPARED); + comm, PERSISTENT | RECV | PREPARED); return request; } @@ -546,6 +590,8 @@ void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm); smpi_mpi_startall(2, requests); smpi_mpi_waitall(2, requests, stats); + smpi_mpi_request_free(&requests[0]); + smpi_mpi_request_free(&requests[1]); if(status != MPI_STATUS_IGNORE) { // Copy receive status *status = stats[1]; @@ -562,7 +608,8 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) MPI_Request req = *request; smpi_empty_status(status); - if(!(req->detached && req->flags & SEND) && !(req->flags & PREPARED)){ + if(!(req->detached && req->flags & SEND) + && !(req->flags & PREPARED)){ if(status != MPI_STATUS_IGNORE) { int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src; status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src); @@ -577,14 +624,28 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) 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->flags & ACCUMULATE) || (datatype->has_subtype == 1)){ + if (!_xbt_replay_is_active()){ + if( smpi_privatize_global_variables + && ((char*)req->old_buf >= start_data_exe) + && ((char*)req->old_buf < start_data_exe + size_data_exe ) + ){ + XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment "); + switch_data_segment(smpi_process_index()); + } + } + + 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, req->op); + if(req->detached == 0) free(req->buf); + }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate + int n =req->real_size/smpi_datatype_size(datatype); + smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype); } - if(req->detached == 0) free(req->buf); } smpi_comm_unuse(req->comm); smpi_datatype_unuse(datatype); @@ -604,12 +665,12 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) 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->flags & PERSISTENT) req->action = NULL; - } + req->flags |= FINISHED; + + smpi_mpi_request_free(request); + } int smpi_mpi_test(MPI_Request * request, MPI_Status * status) { @@ -623,6 +684,7 @@ int smpi_mpi_test(MPI_Request * request, MPI_Status * status) { flag = simcall_comm_test((*request)->action); if (flag) { finish_wait(request, status); + if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT)) *request = MPI_REQUEST_NULL; } } @@ -655,6 +717,7 @@ int smpi_mpi_testany(int count, MPI_Request requests[], int *index, if(i != -1) { *index = map[i]; finish_wait(&requests[*index], status); + if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT)) requests[*index] = MPI_REQUEST_NULL; flag = 1; } @@ -707,7 +770,7 @@ void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){ 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 == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, - comm, NON_PERSISTENT | RECV); + comm, PERSISTENT | RECV); //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls double sleeptime= sg_cfg_get_double("smpi/iprobe"); @@ -762,15 +825,15 @@ void smpi_mpi_wait(MPI_Request * request, MPI_Status * status) if ((*request)->action != NULL) { // this is not a detached send simcall_comm_wait((*request)->action, -1.0); - } - #ifdef HAVE_MC - if(MC_is_active()) + if(MC_is_active() && (*request)->action) (*request)->action->comm.dst_data = NULL; // dangling pointer : dst_data is freed with a wait, need to set it to NULL for system state comparison #endif + } finish_wait(request, status); - *request = MPI_REQUEST_NULL; + if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT)) + *request = MPI_REQUEST_NULL; // FIXME for a detached send, finish_wait is not called: } @@ -789,7 +852,9 @@ int smpi_mpi_waitany(int count, MPI_Request requests[], size = 0; XBT_DEBUG("Wait for one of %d", count); for(i = 0; i < count; i++) { - if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) { + if (requests[i] != MPI_REQUEST_NULL + && !(requests[i]->flags & PREPARED) + && !(requests[i]->flags & FINISHED)) { if (requests[i]->action != NULL) { XBT_DEBUG("Waiting any %p ", requests[i]); xbt_dynar_push(comms, &requests[i]->action); @@ -800,6 +865,7 @@ int smpi_mpi_waitany(int count, MPI_Request requests[], 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 + if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT)) requests[i]=MPI_REQUEST_NULL;//set to null break; } @@ -812,6 +878,7 @@ int smpi_mpi_waitany(int count, MPI_Request requests[], if (i != -1) { index = map[i]; finish_wait(&requests[index], status); + if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT)) requests[index] = MPI_REQUEST_NULL; } } @@ -845,6 +912,7 @@ 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); index = c; @@ -852,6 +920,7 @@ int smpi_mpi_waitall(int count, MPI_Request requests[], index = smpi_mpi_waitany(count, requests, pstat); if (index == MPI_UNDEFINED) break; + if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT)) requests[index]=MPI_REQUEST_NULL; } if (status != MPI_STATUSES_IGNORE) { @@ -881,6 +950,7 @@ int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, if(status != MPI_STATUSES_IGNORE) { status[index] = *pstat; } + if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT)) requests[index]=MPI_REQUEST_NULL; }else{ return MPI_UNDEFINED; @@ -906,8 +976,8 @@ int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, if(status != MPI_STATUSES_IGNORE) { status[i] = *pstat; } + if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT) requests[i]=MPI_REQUEST_NULL; - } }else{ count_dead++; @@ -964,6 +1034,9 @@ void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype, // Wait for completion of irecv's. smpi_mpi_startall(size - 1, requests); smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + for(src = 0; src < size-1; src++) { + smpi_mpi_request_free(&requests[src]); + } xbt_free(requests); } } @@ -1028,6 +1101,9 @@ void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, // Wait for completion of irecv's. smpi_mpi_startall(size - 1, requests); smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + for(src = 0; src < size-1; src++) { + smpi_mpi_request_free(&requests[src]); + } xbt_free(requests); } } @@ -1068,6 +1144,9 @@ void smpi_mpi_allgather(void *sendbuf, int sendcount, // Wait for completion of all comms. smpi_mpi_startall(2 * (size - 1), requests); smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE); + for(other = 0; other < 2*(size-1); other++) { + smpi_mpi_request_free(&requests[other]); + } xbt_free(requests); } @@ -1107,6 +1186,9 @@ void smpi_mpi_allgatherv(void *sendbuf, int sendcount, // Wait for completion of all comms. smpi_mpi_startall(2 * (size - 1), requests); smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE); + for(other = 0; other < 2*(size-1); other++) { + smpi_mpi_request_free(&requests[other]); + } xbt_free(requests); } @@ -1147,6 +1229,9 @@ void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype, // Wait for completion of isend's. smpi_mpi_startall(size - 1, requests); smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + for(dst = 0; dst < size-1; dst++) { + smpi_mpi_request_free(&requests[dst]); + } xbt_free(requests); } } @@ -1188,6 +1273,9 @@ void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, // Wait for completion of isend's. smpi_mpi_startall(size - 1, requests); smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + for(dst = 0; dst < size-1; dst++) { + smpi_mpi_request_free(&requests[dst]); + } xbt_free(requests); } } @@ -1250,6 +1338,8 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, XBT_DEBUG("finished waiting any request with index %d", index); if(index == MPI_UNDEFINED) { break; + }else{ + smpi_mpi_request_free(&requests[index]); } if(op) /* op can be MPI_OP_NULL that does nothing */ smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); @@ -1335,6 +1425,9 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, for(index = 0; index < rank; index++) { xbt_free(tmpbufs[index]); } + for(index = 0; index < size-1; index++) { + smpi_mpi_request_free(&requests[index]); + } xbt_free(tmpbufs); xbt_free(requests); } @@ -1404,6 +1497,9 @@ void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, for(index = 0; index < rank; index++) { xbt_free(tmpbufs[index]); } + for(index = 0; index < size-1; index++) { + smpi_mpi_request_free(&requests[index]); + } xbt_free(tmpbufs); xbt_free(requests); }