X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/5a99bd94a2cedf8b455e78de53b19bebab8c034e..78d66ee7979cae24d0b789adb83a2d1cfebb7bea:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index 57ae7e0b0f..3e10693c7d 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 @@ -7,6 +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" @@ -134,12 +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); @@ -156,12 +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); @@ -178,12 +193,20 @@ 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); @@ -195,7 +218,12 @@ 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 +278,8 @@ 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; @@ -270,35 +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"); - if (action_fp == NULL) - xbt_die("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; @@ -344,17 +341,26 @@ void smpi_mpi_start(MPI_Request request) if (request->flags & RECV) { print_request("New recv", request); - - if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")){ + + int async_small_thres = sg_cfg_get_int("smpi/async_small_thres"); + + xbt_mutex_t mut = smpi_process_mailboxes_mutex(); + if (async_small_thres != 0 ||request->flags & RMA) + xbt_mutex_acquire(mut); + + if (async_small_thres == 0 && !request->flags & RMA) { + mailbox = smpi_process_mailbox(); + } + else if (request->flags & RMA || request->size < 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 in the small mailbox %p (in case of SSEND)?", mailbox); - smx_synchro_t action = SIMIX_comm_iprobe(SIMIX_process_self(), 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){ mailbox = smpi_process_mailbox(); XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox); - action = SIMIX_comm_iprobe(SIMIX_process_self(), mailbox, 0, request->src,request->tag, &match_recv, (void*)request); + 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(); @@ -365,7 +371,7 @@ 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_synchro_t action = SIMIX_comm_iprobe(SIMIX_process_self(), 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"); @@ -386,25 +392,24 @@ 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->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"); - + if (async_small_thres != 0 || request->flags & RMA) + 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 @@ -418,11 +423,21 @@ void smpi_mpi_start(MPI_Request request) simcall_process_sleep(sleeptime); XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size)); } - - if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode + + int async_small_thres = sg_cfg_get_int("smpi/async_small_thres"); + + xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver); + + if (async_small_thres != 0 ||request->flags & RMA) + xbt_mutex_acquire(mut); + + if (!(async_small_thres != 0 ||request->flags & RMA)) { + mailbox = smpi_process_remote_mailbox(receiver); + } + else if (request->flags & RMA || request->size < 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_synchro_t action = SIMIX_comm_iprobe(SIMIX_process_self(), 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); @@ -430,7 +445,7 @@ void smpi_mpi_start(MPI_Request request) } 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 = SIMIX_comm_iprobe(SIMIX_process_self(), mailbox, 1,request->dst, request->tag, &match_send, (void*)request); + 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); @@ -454,11 +469,11 @@ void smpi_mpi_start(MPI_Request request) oldbuf = request->buf; 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); @@ -484,13 +499,12 @@ void smpi_mpi_start(MPI_Request request) -#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_thres != 0 || request->flags & RMA) + xbt_mutex_release(mut); } } @@ -694,8 +708,8 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){ 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()); @@ -719,7 +733,6 @@ 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(); @@ -727,7 +740,6 @@ 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)); @@ -745,8 +757,12 @@ 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); @@ -855,8 +871,9 @@ 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 @@ -864,31 +881,31 @@ 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);