X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/8d5e9d3ae9f7e007e1dfefc1a335d72623335e01..e9f0018b823e34405847177b25a85d3facc30ae1:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index 1887ec0067..497f383505 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -1,4 +1,4 @@ -/* Copyright (c) 2007-2014. The SimGrid Team. +/* Copyright (c) 2007-2015. The SimGrid Team. * All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it @@ -7,9 +7,10 @@ #include "private.h" #include "xbt/virtu.h" #include "mc/mc.h" +#include "src/mc/mc_replay.h" #include "xbt/replay.h" #include -#include "simix/smx_private.h" +#include "src/simix/smx_private.h" #include "surf/surf.h" #include "simgrid/sg_config.h" #include "colls/colls.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); @@ -71,6 +72,11 @@ 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 ?) @@ -129,12 +135,19 @@ static double smpi_os(double size) unsigned int iter = 0; s_smpi_factor_t fact; double current=0.0; + // Iterate over all the sections that were specified and find the right + // value. (fact.factor represents the interval sizes; we want to find the + // section that has fact.factor <= size and no other such fact.factor <= size) + // Note: parse_factor() (used before) already sorts the dynar we iterate over! xbt_dynar_foreach(smpi_os_values, iter, fact) { - if (size <= fact.factor) { + if (size <= fact.factor) { // Values already too large, use the previously + // computed value of current! XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current); return current; }else{ - current=fact.values[0]+fact.values[1]*size; + // If the next section is too large, the current section must be used. + // Hence, save the cost, as we might have to use it. + current = fact.values[0]+fact.values[1]*size; } } XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current); @@ -151,12 +164,19 @@ static double smpi_ois(double size) unsigned int iter = 0; s_smpi_factor_t fact; double current=0.0; + // Iterate over all the sections that were specified and find the right + // value. (fact.factor represents the interval sizes; we want to find the + // section that has fact.factor <= size and no other such fact.factor <= size) + // Note: parse_factor() (used before) already sorts the dynar we iterate over! xbt_dynar_foreach(smpi_ois_values, iter, fact) { - if (size <= fact.factor) { + if (size <= fact.factor) { // Values already too large, use the previously + // computed value of current! XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current); return current; }else{ - current=fact.values[0]+fact.values[1]*size; + // If the next section is too large, the current section must be used. + // Hence, save the cost, as we might have to use it. + current = fact.values[0]+fact.values[1]*size; } } XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current); @@ -173,18 +193,45 @@ static double smpi_or(double size) unsigned int iter = 0; s_smpi_factor_t fact; double current=0.0; + // Iterate over all the sections that were specified and find the right + // value. (fact.factor represents the interval sizes; we want to find the + // section that has fact.factor <= size and no other such fact.factor <= size) + // Note: parse_factor() (used before) already sorts the dynar we iterate over! xbt_dynar_foreach(smpi_or_values, iter, fact) { - if (size <= fact.factor) { + if (size <= fact.factor) { // Values already too large, use the previously + // computed value of current! XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current); return current; - }else + } else { + // If the next section is too large, the current section must be used. + // Hence, save the cost, as we might have to use it. current=fact.values[0]+fact.values[1]*size; + } } XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current); 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 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 { + 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 +244,11 @@ 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); } } @@ -226,12 +273,13 @@ static MPI_Request build_request(void *buf, int count, request->truncated = 0; request->real_size = 0; request->real_tag = 0; - - request->refcount=1; -#ifdef HAVE_TRACING + if(flags & PERSISTENT) + request->refcount = 1; + else + request->refcount = 0; + request->op = MPI_REPLACE; request->send = 0; request->recv = 0; -#endif if (flags & SEND) smpi_datatype_unuse(datatype); return request; @@ -248,35 +296,6 @@ void smpi_empty_status(MPI_Status * status) } } -void smpi_action_trace_run(char *path) -{ - char *name; - xbt_dynar_t todo; - xbt_dict_cursor_t cursor; - - action_fp=NULL; - if (path) { - action_fp = fopen(path, "r"); - xbt_assert(action_fp != NULL, "Cannot open %s: %s", path, - strerror(errno)); - } - - if (!xbt_dict_is_empty(action_queues)) { - XBT_WARN - ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file."); - - - xbt_dict_foreach(action_queues, cursor, name, todo) { - XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name); - } - } - - if (path) - fclose(action_fp); - xbt_dict_free(&action_queues); - action_queues = xbt_dict_new_homogeneous(NULL); -} - static void smpi_mpi_request_free_voidp(void* request) { MPI_Request req = request; @@ -290,7 +309,6 @@ 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 +318,6 @@ 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; } @@ -310,107 +327,184 @@ 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; } void smpi_mpi_start(MPI_Request request) { smx_rdv_t mailbox; - + 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")) - mailbox = smpi_process_mailbox_small(); - else + + int async_small_thresh = sg_cfg_get_int("smpi/async_small_thresh"); + + xbt_mutex_t mut = smpi_process_mailboxes_mutex(); + if (async_small_thresh != 0 ||request->flags & RMA) + xbt_mutex_acquire(mut); + + if (async_small_thresh == 0 && !(request->flags & RMA)) { mailbox = smpi_process_mailbox(); + } + else if (request->flags & RMA || request->size < async_small_thresh){ + //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 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 nothing, switch back to the small mailbox : %p", mailbox); + mailbox = smpi_process_mailbox_small(); + } + }else{ + XBT_DEBUG("yes there was something for us in the large mailbox"); + } + }else{ + mailbox = smpi_process_mailbox_small(); + XBT_DEBUG("Is there a corresponding send already posted the small mailbox?"); + 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"); + mailbox = smpi_process_mailbox(); + }else{ + XBT_DEBUG("yes there was something for us in the small mailbox"); + } + } + + //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->action = simcall_comm_irecv(SIMIX_process_self(), mailbox, request->buf, &request->real_size, &match_recv, + !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){ + if (async_small_thresh != 0 || request->flags & RMA) + xbt_mutex_release(mut); + } else { + + + int receiver = request->dst; + + 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("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size)); + XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size)); } - } else { + int async_small_thresh = sg_cfg_get_int("smpi/async_small_thresh"); + xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver); - int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst); + if (async_small_thresh != 0 || request->flags & RMA) + xbt_mutex_acquire(mut); - #ifdef HAVE_TRACING - int rank = smpi_process_index(); - if (TRACE_smpi_view_internals()) { - TRACE_smpi_send(rank, rank, receiver,request->size); + if (!(async_small_thresh != 0 || request->flags & RMA)) { + mailbox = smpi_process_remote_mailbox(receiver); + } + else if (request->flags & RMA || request->size < async_small_thresh) { // eager mode + mailbox = smpi_process_remote_mailbox(receiver); + XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox); + 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); + XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox); + } else{ + mailbox = smpi_process_remote_mailbox_small(receiver); + XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox); + action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request); + if(action ==NULL){ + XBT_DEBUG("No, we are first, send to large mailbox"); + mailbox = smpi_process_remote_mailbox(receiver); + } + } + }else{ + XBT_DEBUG("Yes there was something for us in the large mailbox"); } - #endif -/* if(receiver == MPI_UNDEFINED) {*/ -/* XBT_WARN("Trying to send a message to a wrong rank");*/ -/* return;*/ -/* }*/ - print_request("New send", request); - if (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); + XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf); } + + 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 (!smpi_process_get_replaying() && oldbuf && request->size!=0){ + if((smpi_privatize_global_variables) + && ((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); + } + 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 request->real_size=request->size; smpi_datatype_use(request->old_type); 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; - 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)); - } - 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_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"); + + -#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()); + if (async_small_thresh != 0 || request->flags & RMA) + xbt_mutex_release(mut); } } @@ -437,19 +531,50 @@ 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; } @@ -473,14 +598,12 @@ 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; } @@ -548,6 +671,8 @@ 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]; @@ -564,7 +689,8 @@ 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); @@ -579,21 +705,34 @@ 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 (!smpi_process_get_replaying()){ + if( smpi_privatize_global_variables + && ((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()); + } + } + + 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); } -#ifdef HAVE_TRACING if (TRACE_smpi_view_internals()) { if(req->flags & RECV){ int rank = smpi_process_index(); @@ -601,23 +740,32 @@ 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)); } - - 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) { int flag; //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before) + + // 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); + smpi_empty_status(status); flag = 1; if (!((*request)->flags & PREPARED)) { @@ -625,7 +773,11 @@ 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; @@ -640,7 +792,7 @@ 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++) { @@ -652,13 +804,21 @@ 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) { *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; + nsleeps=1; + }else{ + nsleeps++; } }else{ //all requests are null or inactive, return true @@ -709,45 +869,43 @@ 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"); - //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; - - 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; 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, 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 (sg_cfg_get_int("smpi/async_small_thresh")>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); @@ -764,14 +922,14 @@ 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() || 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); - *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: } @@ -785,12 +943,14 @@ 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); 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); @@ -801,6 +961,7 @@ 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; } @@ -813,6 +974,7 @@ 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; } } @@ -846,13 +1008,15 @@ 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 { 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) { @@ -882,6 +1046,7 @@ 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; @@ -907,8 +1072,8 @@ 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++; @@ -921,14 +1086,14 @@ int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm) { - // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI)) - nary_tree_bcast(buf, count, datatype, root, comm, 4); + // use openMPI selector as default one, as naive one was removed + smpi_coll_tuned_bcast_ompi(buf, count, datatype, root, comm); } void smpi_mpi_barrier(MPI_Comm comm) { - // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI)) - nary_tree_barrier(comm, 4); + // use openMPI selector as default one, as naive one was removed + smpi_coll_tuned_barrier_ompi(comm); } void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype, @@ -965,6 +1130,9 @@ 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); } } @@ -986,12 +1154,13 @@ 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, @@ -1029,6 +1198,9 @@ 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); } } @@ -1069,6 +1241,9 @@ 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); } @@ -1108,6 +1283,9 @@ 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); } @@ -1148,6 +1326,9 @@ 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); } } @@ -1189,6 +1370,9 @@ 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); } } @@ -1206,7 +1390,7 @@ 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); } @@ -1237,7 +1421,10 @@ 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); @@ -1251,18 +1438,20 @@ 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); } - 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); } } } @@ -1299,7 +1488,7 @@ 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); @@ -1334,7 +1523,10 @@ 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]); } xbt_free(tmpbufs); xbt_free(requests); @@ -1362,7 +1554,7 @@ 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); @@ -1403,7 +1595,10 @@ 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]); } xbt_free(tmpbufs); xbt_free(requests);