X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/0dc10aaf6accd03c5b68e8ff178f6aa843e84c85..e7ca102c2884a457fb10de086fbaf074506c02e1:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index 2bb5b8c15c..3ea6ac248c 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -71,6 +71,11 @@ xbt_dynar_t smpi_os_values = NULL; xbt_dynar_t smpi_or_values = NULL; xbt_dynar_t smpi_ois_values = NULL; +double smpi_wtime_sleep = 0.0; +double smpi_iprobe_sleep = 1e-4; +double smpi_test_sleep = 1e-4; + + // Methods used to parse and store the values for timing injections in smpi // These are taken from surf/network.c and generalized to have more factors // These methods should be merged with those in surf/network.c (moved somewhere in xbt ?) @@ -185,6 +190,20 @@ static double smpi_or(double size) return current; } +double smpi_mpi_wtime(){ + double time; + 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 + if(smpi_wtime_sleep > 0) simcall_process_sleep(smpi_wtime_sleep); + smpi_bench_begin(); + } else { + time = SIMIX_get_clock(); + } + return time; +} + static MPI_Request build_request(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags) @@ -197,11 +216,11 @@ 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); } } @@ -230,7 +249,7 @@ static MPI_Request build_request(void *buf, int count, request->refcount = 1; else request->refcount = 0; - + request->op = MPI_REPLACE; #ifdef HAVE_TRACING request->send = 0; request->recv = 0; @@ -260,8 +279,8 @@ 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)) { @@ -326,7 +345,7 @@ void smpi_mpi_start(MPI_Request request) 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(); @@ -336,6 +355,7 @@ void smpi_mpi_start(MPI_Request request) smpi_comm_use(request->comm); 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 @@ -351,7 +371,7 @@ 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); } @@ -361,24 +381,33 @@ 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 (!_xbt_replay_is_active() && oldbuf && request->size!=0){ - request->buf = xbt_malloc(request->size); - memcpy(request->buf,oldbuf,request->size); + 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 @@ -387,22 +416,22 @@ 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, - &xbt_free, // 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); @@ -446,6 +475,38 @@ void smpi_mpi_request_free(MPI_Request * 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) { @@ -475,8 +536,6 @@ 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) { @@ -584,14 +643,28 @@ 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); @@ -623,6 +696,12 @@ 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) + + //to avoid deadlocks + //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); + smpi_empty_status(status); flag = 1; if (!((*request)->flags & PREPARED)) { @@ -630,8 +709,11 @@ int smpi_mpi_test(MPI_Request * request, MPI_Status * status) { flag = simcall_comm_test((*request)->action); if (flag) { finish_wait(request, status); + nsleeps=1;//reset the number of sleeps we will do next time if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT)) *request = MPI_REQUEST_NULL; + }else{ + nsleeps++; } } return flag; @@ -658,6 +740,10 @@ int smpi_mpi_testany(int count, MPI_Request requests[], int *index, } } if(size > 0) { + //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it + static int nsleeps = 1; + if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep); + i = simcall_comm_testany(comms); // not MPI_UNDEFINED, as this is a simix return code if(i != -1) { @@ -666,6 +752,9 @@ int smpi_mpi_testany(int count, MPI_Request requests[], int *index, if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT)) requests[*index] = MPI_REQUEST_NULL; flag = 1; + nsleeps=1; + }else{ + nsleeps++; } }else{ //all requests are null or inactive, return true @@ -719,12 +808,9 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* 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"); //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it static int nsleeps = 1; - - simcall_process_sleep(sleeptime); - + if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep); // behave like a receive, but don't do it smx_rdv_t mailbox; @@ -772,7 +858,7 @@ 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 }