X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/87222536f2f0f6fa81d0057c6ab924ede70d9ed2..55eda6695068477d25d74cfa9335f79ad5a9b826:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index b58dd244a1..84edb27fe1 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -7,6 +7,7 @@ #include "private.h" #include "xbt/virtu.h" #include "mc/mc.h" +#include "mc/mc_replay.h" #include "xbt/replay.h" #include #include "simix/smx_private.h" @@ -17,7 +18,7 @@ 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) { +static int match_recv(void* a, void* b, smx_synchro_t ignored) { 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); @@ -39,7 +40,7 @@ static int match_recv(void* a, void* b, smx_action_t ignored) { }else return 0; } -static int match_send(void* a, void* b,smx_action_t ignored) { +static int match_send(void* a, void* b,smx_synchro_t ignored) { 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); @@ -195,7 +196,12 @@ double smpi_mpi_wtime(){ if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_process_get_sampling()) { smpi_bench_end(); time = SIMIX_get_clock(); - //to avoid deadlocks if called too many times + // to avoid deadlocks if used as a break condition, such as + // while (MPI_Wtime(...) < time_limit) { + // .... + // } + // because the time will not normally advance when only calls to MPI_Wtime + // are made -> deadlock (MPI_Wtime never reaches the time limit) if(smpi_wtime_sleep > 0) simcall_process_sleep(smpi_wtime_sleep); smpi_bench_begin(); } else { @@ -250,10 +256,8 @@ static MPI_Request build_request(void *buf, int count, else request->refcount = 0; request->op = MPI_REPLACE; -#ifdef HAVE_TRACING request->send = 0; request->recv = 0; -#endif if (flags & SEND) smpi_datatype_unuse(datatype); return request; @@ -344,19 +348,22 @@ void smpi_mpi_start(MPI_Request request) if (request->flags & RECV) { print_request("New recv", request); + + xbt_mutex_t mut=smpi_process_mailboxes_mutex(); + xbt_mutex_acquire(mut); if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")){ //We have to check both mailboxes (because SSEND messages are sent to the large mbox). begin with the more appropriate one : the small one. mailbox = smpi_process_mailbox_small(); - XBT_DEBUG("Is there a corresponding send already posted the small mailbox %p (in case of SSEND)?", mailbox); - smx_action_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); + XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox); + smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); if(action ==NULL){ mailbox = smpi_process_mailbox(); XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox); action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); if(action ==NULL){ - XBT_DEBUG("Still notching, switch back to the small mailbox : %p", mailbox); + XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox); mailbox = smpi_process_mailbox_small(); } }else{ @@ -365,7 +372,7 @@ void smpi_mpi_start(MPI_Request request) }else{ mailbox = smpi_process_mailbox_small(); XBT_DEBUG("Is there a corresponding send already posted the small mailbox?"); - smx_action_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); + smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); if(action ==NULL){ XBT_DEBUG("No, nothing in the permanent receive mailbox"); @@ -375,38 +382,55 @@ void smpi_mpi_start(MPI_Request request) } } + //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; + if(sleeptime!=0.0){ + simcall_process_sleep(sleeptime); + XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size)); + } + // 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; smpi_datatype_use(request->old_type); smpi_comm_use(request->comm); request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, - &smpi_comm_copy_buffer_callback, + !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback + : &smpi_comm_null_copy_buffer_callback, request, -1.0); XBT_DEBUG("recv simcall posted"); - //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; - if(sleeptime!=0.0){ - simcall_process_sleep(sleeptime); - XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size)); - } + xbt_mutex_release(mut); } else { int receiver = request->dst; - #ifdef HAVE_TRACING - int rank = request->src; - if (TRACE_smpi_view_internals()) { - TRACE_smpi_send(rank, rank, receiver,request->size); - } - #endif + int rank = request->src; + if (TRACE_smpi_view_internals()) { + TRACE_smpi_send(rank, rank, receiver,request->size); + } print_request("New send", request); + + //if we are giving back the control to the user without waiting for completion, we have to inject timings + 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){ + simcall_process_sleep(sleeptime); + XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size)); + } + + xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver); + xbt_mutex_acquire(mut); + if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode mailbox = smpi_process_remote_mailbox(receiver); XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox); - smx_action_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request); + smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request); if(action ==NULL){ if (! (request->flags & SSEND)){ mailbox = smpi_process_remote_mailbox_small(receiver); @@ -436,13 +460,13 @@ void smpi_mpi_start(MPI_Request request) request->refcount++; if(request->old_type->has_subtype == 0){ oldbuf = request->buf; - if (!_xbt_replay_is_active() && oldbuf && request->size!=0){ + if (!smpi_process_get_replaying() && oldbuf && request->size!=0){ if((smpi_privatize_global_variables) - && ((char*)request->buf >= start_data_exe) - && ((char*)request->buf < start_data_exe + size_data_exe )){ + && ((char*) request->buf >= smpi_start_data_exe) + && ((char*)request->buf < smpi_start_data_exe + smpi_size_data_exe )){ XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment "); - smpi_switch_data_segment(request->src); - } + smpi_switch_data_segment(request->src); + } buf = xbt_malloc(request->size); memcpy(buf,oldbuf,request->size); XBT_DEBUG("buf %p copied into %p",oldbuf,buf); @@ -459,31 +483,20 @@ void smpi_mpi_start(MPI_Request request) buf, request->real_size, &match_send, &xbt_free_f, // how to free the userdata if a detached send fails - &smpi_comm_copy_buffer_callback, + !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback + : &smpi_comm_null_copy_buffer_callback, request, // detach if msg size < eager/rdv switch limit request->detached); XBT_DEBUG("send simcall posted"); - //if we are giving back the control to the user without waiting for completion, we have to inject timings - 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){ - simcall_process_sleep(sleeptime); - XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size)); - } -#ifdef HAVE_TRACING /* FIXME: detached sends are not traceable (request->action == NULL) */ if (request->action) - simcall_set_category(request->action, TRACE_internal_smpi_get_category()); - -#endif + simcall_set_category(request->action, TRACE_internal_smpi_get_category()); + xbt_mutex_release(mut); } } @@ -685,10 +698,10 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) MPI_Datatype datatype = req->old_type; if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){ - if (!_xbt_replay_is_active()){ + if (!smpi_process_get_replaying()){ if( smpi_privatize_global_variables - && ((char*)req->old_buf >= start_data_exe) - && ((char*)req->old_buf < start_data_exe + size_data_exe ) + && ((char*)req->old_buf >= smpi_start_data_exe) + && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe ) ){ XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment "); smpi_switch_data_segment(smpi_process_index()); @@ -712,7 +725,6 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) } -#ifdef HAVE_TRACING if (TRACE_smpi_view_internals()) { if(req->flags & RECV){ int rank = smpi_process_index(); @@ -720,7 +732,6 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) TRACE_smpi_recv(rank, src_traced, rank); } } -#endif if(req->detached_sender!=NULL){ smpi_mpi_request_free(&(req->detached_sender)); @@ -738,8 +749,12 @@ int smpi_mpi_test(MPI_Request * request, MPI_Status * status) { //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before) - //to avoid deadlocks - //multiplier to the sleeptime, to increase speed of execution, each failed test will increase it + // to avoid deadlocks if used as a break condition, such as + // while (MPI_Test(request, flag, status) && flag) { + // } + // because the time will not normally advance when only calls to MPI_Test + // are made -> deadlock + // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it static int nsleeps = 1; if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep); @@ -769,7 +784,7 @@ int smpi_mpi_testany(int count, MPI_Request requests[], int *index, *index = MPI_UNDEFINED; flag = 0; - comms = xbt_dynar_new(sizeof(smx_action_t), NULL); + comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL); map = xbt_new(int, count); size = 0; for(i = 0; i < count; i++) { @@ -848,8 +863,9 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_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, 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 - //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it + // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls + // (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... ) + // multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it static int nsleeps = 1; if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep); // behave like a receive, but don't do it @@ -857,31 +873,31 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* print_request("New iprobe", request); // We have to test both mailboxes as we don't know if we will receive one one or another - if (sg_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, 0, 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, 0, request->src,request->tag, &match_recv, (void*)request); - } + if (sg_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, 0, 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, 0, request->src,request->tag, &match_recv, (void*)request); + } - if(request->action){ + if (request->action){ MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action); *flag = 1; if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) { status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src); - status->MPI_TAG = req->tag; - status->MPI_ERROR = MPI_SUCCESS; - status->count = req->real_size; + status->MPI_TAG = req->tag; + status->MPI_ERROR = MPI_SUCCESS; + status->count = req->real_size; } nsleeps=1;//reset the number of sleeps we will do next time } else { - *flag = 0; - nsleeps++; + *flag = 0; + nsleeps++; } smpi_mpi_request_free(&request); @@ -898,10 +914,9 @@ 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() && (*request)->action) + + if((MC_is_active() || MC_record_replay_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); @@ -920,7 +935,7 @@ int smpi_mpi_waitany(int count, MPI_Request requests[], index = MPI_UNDEFINED; if(count > 0) { // Wait for a request to complete - comms = xbt_dynar_new(sizeof(smx_action_t), NULL); + comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL); map = xbt_new(int, count); size = 0; XBT_DEBUG("Wait for one of %d", count); @@ -986,7 +1001,7 @@ int smpi_mpi_waitall(int count, MPI_Request requests[], } for(c = 0; c < count; c++) { - if (MC_is_active()) { + if (MC_is_active() || MC_record_replay_is_active()) { smpi_mpi_wait(&requests[c], pstat); index = c; } else { @@ -1131,12 +1146,13 @@ void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, displs[i] = count; count += recvcounts[i]; } - tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype)); + tmpbuf=(void*)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)); + mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm); smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf, recvcounts[rank], datatype, 0, comm); xbt_free(displs); - xbt_free(tmpbuf); + smpi_free_tmp_buffer(tmpbuf); } void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, @@ -1366,7 +1382,7 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, char* sendtmpbuf = (char*) sendbuf; if( sendbuf == MPI_IN_PLACE ) { - sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype)); + sendtmpbuf = (char *)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)); smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype); } @@ -1397,7 +1413,10 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, if(src != root) { // FIXME: possibly overkill we we have contiguous/noncontiguous data // mapping... - tmpbufs[index] = xbt_malloc(count * dataext); + if (!smpi_process_get_replaying()) + tmpbufs[index] = xbt_malloc(count * dataext); + else + tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext); requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, src, system_tag, comm); @@ -1417,14 +1436,14 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, if(op) /* op can be MPI_OP_NULL that does nothing */ smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); } - for(index = 0; index < size - 1; index++) { - xbt_free(tmpbufs[index]); - } + for(index = 0; index < size - 1; index++) { + smpi_free_tmp_buffer(tmpbufs[index]); + } xbt_free(tmpbufs); xbt_free(requests); if( sendbuf == MPI_IN_PLACE ) { - xbt_free(sendtmpbuf); + smpi_free_tmp_buffer(sendtmpbuf); } } } @@ -1461,7 +1480,7 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, for(other = 0; other < rank; other++) { // FIXME: possibly overkill we we have contiguous/noncontiguous data // mapping... - tmpbufs[index] = xbt_malloc(count * dataext); + tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext); requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm); @@ -1496,7 +1515,7 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, } } for(index = 0; index < rank; index++) { - xbt_free(tmpbufs[index]); + smpi_free_tmp_buffer(tmpbufs[index]); } for(index = 0; index < size-1; index++) { smpi_mpi_request_free(&requests[index]); @@ -1527,7 +1546,7 @@ void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, for(other = 0; other < rank; other++) { // FIXME: possibly overkill we we have contiguous/noncontiguous data // mapping... - tmpbufs[index] = xbt_malloc(count * dataext); + tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext); requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm); @@ -1568,7 +1587,7 @@ void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, } } for(index = 0; index < rank; index++) { - xbt_free(tmpbufs[index]); + smpi_free_tmp_buffer(tmpbufs[index]); } for(index = 0; index < size-1; index++) { smpi_mpi_request_free(&requests[index]);