X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/d0ffc37686edcc803601f76ab51fdfed5fc2f241..60449c1d78a05619fe3d1a77fa1d7d0c20a79dba:/src/smpi/smpi_replay.c diff --git a/src/smpi/smpi_replay.c b/src/smpi/smpi_replay.c index fbfb10709d..f08faf1cc6 100644 --- a/src/smpi/smpi_replay.c +++ b/src/smpi/smpi_replay.c @@ -1,10 +1,9 @@ -/* Copyright (c) 2009-2013. The SimGrid Team. +/* Copyright (c) 2009-2014. The SimGrid Team. * All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ - #include "private.h" #include #include @@ -19,6 +18,11 @@ xbt_dynar_t *reqq = NULL; MPI_Datatype MPI_DEFAULT_TYPE; MPI_Datatype MPI_CURRENT_TYPE; +static int sendbuffer_size=0; +char* sendbuffer=NULL; +static int recvbuffer_size=0; +char* recvbuffer=NULL; + static void log_timed_action (const char *const *action, double clock){ if (XBT_LOG_ISENABLED(smpi_replay, xbt_log_priority_verbose)){ char *name = xbt_str_join_array(action, " "); @@ -27,10 +31,31 @@ static void log_timed_action (const char *const *action, double clock){ } } -typedef struct { - xbt_dynar_t irecvs; /* of MPI_Request */ -} s_smpi_replay_globals_t, *smpi_replay_globals_t; +//allocate a single buffer for all sends, growing it if needed +void* smpi_get_tmp_sendbuffer(int size){ + if (!smpi_process_get_replaying()) + return xbt_malloc(size); + if (sendbuffer_sizeirecvs = xbt_dynar_new(sizeof(MPI_Request),NULL); - + CHECK_ACTION_PARAMS(action, 0, 1); if(action[2]) MPI_DEFAULT_TYPE= MPI_DOUBLE; // default MPE dataype else MPI_DEFAULT_TYPE= MPI_BYTE; // default TAU datatype - smpi_process_set_user_data((void*) globals); - /* start a simulated timer */ smpi_process_simulated_start(); /*initialize the number of active processes */ @@ -98,21 +160,13 @@ static void action_init(const char *const *action) reqq=xbt_new0(xbt_dynar_t,active_processes); for(i=0;iirecvs)); - xbt_dynar_free_container(&(globals->irecvs)); - } - free(globals); } static void action_comm_size(const char *const *action) @@ -139,14 +193,27 @@ static void action_comm_dup(const char *const *action) static void action_compute(const char *const *action) { + CHECK_ACTION_PARAMS(action, 1, 0); double clock = smpi_process_simulated_elapsed(); - smpi_execute_flops(parse_double(action[2])); + double flops= parse_double(action[2]); +#ifdef HAVE_TRACING + int rank = smpi_process_index(); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type=TRACING_COMPUTING; + extra->comp_size=flops; + TRACE_smpi_computing_in(rank, extra); +#endif + smpi_execute_flops(flops); +#ifdef HAVE_TRACING + TRACE_smpi_computing_out(rank); +#endif log_timed_action (action, clock); } static void action_send(const char *const *action) { + CHECK_ACTION_PARAMS(action, 2, 1); int to = atoi(action[2]); double size=parse_double(action[3]); double clock = smpi_process_simulated_elapsed(); @@ -158,10 +225,16 @@ static void action_send(const char *const *action) } #ifdef HAVE_TRACING - int rank = smpi_comm_rank(MPI_COMM_WORLD); - TRACE_smpi_computing_out(rank); + int rank = smpi_process_index(); + int dst_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), to); - TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, size*smpi_datatype_size(MPI_CURRENT_TYPE)); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_SEND; + extra->send_size = size; + extra->src = rank; + extra->dst = dst_traced; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra); TRACE_smpi_send(rank, rank, dst_traced, size*smpi_datatype_size(MPI_CURRENT_TYPE)); #endif @@ -171,13 +244,13 @@ static void action_send(const char *const *action) #ifdef HAVE_TRACING TRACE_smpi_ptp_out(rank, rank, dst_traced, __FUNCTION__); - TRACE_smpi_computing_in(rank); #endif } static void action_Isend(const char *const *action) { + CHECK_ACTION_PARAMS(action, 2, 1); int to = atoi(action[2]); double size=parse_double(action[3]); double clock = smpi_process_simulated_elapsed(); @@ -187,10 +260,15 @@ static void action_Isend(const char *const *action) else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; #ifdef HAVE_TRACING - int rank = smpi_comm_rank(MPI_COMM_WORLD); - TRACE_smpi_computing_out(rank); + int rank = smpi_process_index(); int dst_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), to); - TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, size*smpi_datatype_size(MPI_CURRENT_TYPE)); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_ISEND; + extra->send_size = size; + extra->src = rank; + extra->dst = dst_traced; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra); TRACE_smpi_send(rank, rank, dst_traced, size*smpi_datatype_size(MPI_CURRENT_TYPE)); #endif @@ -199,15 +277,15 @@ static void action_Isend(const char *const *action) #ifdef HAVE_TRACING TRACE_smpi_ptp_out(rank, rank, dst_traced, __FUNCTION__); request->send = 1; - TRACE_smpi_computing_in(rank); #endif - xbt_dynar_push(reqq[smpi_comm_rank(MPI_COMM_WORLD)],&request); + xbt_dynar_push(reqq[smpi_process_index()],&request); log_timed_action (action, clock); } static void action_recv(const char *const *action) { + CHECK_ACTION_PARAMS(action, 2, 1); int from = atoi(action[2]); double size=parse_double(action[3]); double clock = smpi_process_simulated_elapsed(); @@ -217,19 +295,29 @@ static void action_recv(const char *const *action) { else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; #ifdef HAVE_TRACING - int rank = smpi_comm_rank(MPI_COMM_WORLD); + int rank = smpi_process_index(); int src_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), from); - TRACE_smpi_computing_out(rank); - TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__, size*smpi_datatype_size(MPI_CURRENT_TYPE)); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_RECV; + extra->send_size = size; + extra->src = src_traced; + extra->dst = rank; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__, extra); #endif + //unknow size from the receiver pov + if(size==-1){ + smpi_mpi_probe(from, 0, MPI_COMM_WORLD, &status); + size=status.count; + } + smpi_mpi_recv(NULL, size, MPI_CURRENT_TYPE, from, 0, MPI_COMM_WORLD, &status); #ifdef HAVE_TRACING TRACE_smpi_ptp_out(rank, src_traced, rank, __FUNCTION__); TRACE_smpi_recv(rank, src_traced, rank); - TRACE_smpi_computing_in(rank); #endif log_timed_action (action, clock); @@ -237,22 +325,32 @@ static void action_recv(const char *const *action) { static void action_Irecv(const char *const *action) { + CHECK_ACTION_PARAMS(action, 2, 1); int from = atoi(action[2]); double size=parse_double(action[3]); double clock = smpi_process_simulated_elapsed(); MPI_Request request; - smpi_replay_globals_t globals = - (smpi_replay_globals_t) smpi_process_get_user_data(); - if(action[4]) MPI_CURRENT_TYPE=decode_datatype(action[4]); else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; #ifdef HAVE_TRACING - int rank = smpi_comm_rank(MPI_COMM_WORLD); + int rank = smpi_process_index(); int src_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), from); - TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__, size*smpi_datatype_size(MPI_CURRENT_TYPE)); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_IRECV; + extra->send_size = size; + extra->src = src_traced; + extra->dst = rank; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__, extra); #endif + MPI_Status status; + //unknow size from the receiver pov + if(size==-1){ + smpi_mpi_probe(from, 0, MPI_COMM_WORLD, &status); + size=status.count; + } request = smpi_mpi_irecv(NULL, size, MPI_CURRENT_TYPE, from, 0, MPI_COMM_WORLD); @@ -260,34 +358,75 @@ static void action_Irecv(const char *const *action) TRACE_smpi_ptp_out(rank, src_traced, rank, __FUNCTION__); request->recv = 1; #endif - xbt_dynar_push(globals->irecvs,&request); - xbt_dynar_push(reqq[smpi_comm_rank(MPI_COMM_WORLD)],&request); + xbt_dynar_push(reqq[smpi_process_index()],&request); + + log_timed_action (action, clock); +} + +static void action_test(const char *const *action){ + CHECK_ACTION_PARAMS(action, 0, 0); + double clock = smpi_process_simulated_elapsed(); + MPI_Request request; + MPI_Status status; + int flag = TRUE; + + request = xbt_dynar_pop_as(reqq[smpi_process_index()],MPI_Request); + //if request is null here, this may mean that a previous test has succeeded + //Different times in traced application and replayed version may lead to this + //In this case, ignore the extra calls. + if(request){ +#ifdef HAVE_TRACING + int rank = smpi_process_index(); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type=TRACING_TEST; + TRACE_smpi_testing_in(rank, extra); +#endif + + flag = smpi_mpi_test(&request, &status); + + XBT_DEBUG("MPI_Test result: %d", flag); + /* push back request in dynar to be caught by a subsequent wait. if the test + * did succeed, the request is now NULL. + */ + xbt_dynar_push_as(reqq[smpi_process_index()],MPI_Request, request); +#ifdef HAVE_TRACING + TRACE_smpi_testing_out(rank); +#endif + } log_timed_action (action, clock); } static void action_wait(const char *const *action){ + CHECK_ACTION_PARAMS(action, 0, 0); double clock = smpi_process_simulated_elapsed(); MPI_Request request; MPI_Status status; - smpi_replay_globals_t globals = - (smpi_replay_globals_t) smpi_process_get_user_data(); - xbt_assert(xbt_dynar_length(globals->irecvs), - "action wait not preceded by any irecv: %s", + xbt_assert(xbt_dynar_length(reqq[smpi_process_index()]), + "action wait not preceded by any irecv or isend: %s", xbt_str_join_array(action," ")); - request = xbt_dynar_pop_as(globals->irecvs,MPI_Request); + request = xbt_dynar_pop_as(reqq[smpi_process_index()],MPI_Request); + + if (!request){ + /* Assuming that the trace is well formed, this mean the comm might have + * been caught by a MPI_test. Then just return. + */ + return; + } + #ifdef HAVE_TRACING - int rank = request && request->comm != MPI_COMM_NULL + int rank = request->comm != MPI_COMM_NULL ? smpi_comm_rank(request->comm) : -1; - TRACE_smpi_computing_out(rank); MPI_Group group = smpi_comm_group(request->comm); int src_traced = smpi_group_rank(group, request->src); int dst_traced = smpi_group_rank(group, request->dst); int is_wait_for_receive = request->recv; - TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__, -1); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_WAIT; + TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__, extra); #endif smpi_mpi_wait(&request, &status); #ifdef HAVE_TRACING @@ -295,18 +434,18 @@ static void action_wait(const char *const *action){ if (is_wait_for_receive) { TRACE_smpi_recv(rank, src_traced, dst_traced); } - TRACE_smpi_computing_in(rank); #endif log_timed_action (action, clock); } static void action_waitall(const char *const *action){ + CHECK_ACTION_PARAMS(action, 0, 0); double clock = smpi_process_simulated_elapsed(); int count_requests=0; unsigned int i=0; - count_requests=xbt_dynar_length(reqq[smpi_comm_rank(MPI_COMM_WORLD)]); + count_requests=xbt_dynar_length(reqq[smpi_process_index()]); if (count_requests>0) { MPI_Request requests[count_requests]; @@ -314,7 +453,7 @@ static void action_waitall(const char *const *action){ /* The reqq is an array of dynars. Its index corresponds to the rank. Thus each rank saves its own requests to the array request. */ - xbt_dynar_foreach(reqq[smpi_comm_rank(MPI_COMM_WORLD)],i,requests[i]); + xbt_dynar_foreach(reqq[smpi_process_index()],i,requests[i]); #ifdef HAVE_TRACING //save information from requests @@ -345,10 +484,11 @@ static void action_waitall(const char *const *action){ } } int rank_traced = smpi_process_index(); - TRACE_smpi_computing_out(rank_traced); - - TRACE_smpi_ptp_in(rank_traced, -1, -1, __FUNCTION__, count_requests); - #endif + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_WAITALL; + extra->send_size=count_requests; + TRACE_smpi_ptp_in(rank_traced, -1, -1, __FUNCTION__,extra); + #endif smpi_mpi_waitall(count_requests, requests, status); @@ -367,10 +507,11 @@ static void action_waitall(const char *const *action){ xbt_dynar_free(&srcs); xbt_dynar_free(&dsts); xbt_dynar_free(&recvs); - TRACE_smpi_computing_in(rank_traced); #endif - xbt_dynar_reset(reqq[smpi_comm_rank(MPI_COMM_WORLD)]); + int freedrank=smpi_process_index(); + xbt_dynar_free_container(&(reqq[freedrank])); + reqq[freedrank]=xbt_dynar_new(sizeof(MPI_Request),&xbt_free_ref); } log_timed_action (action, clock); } @@ -378,14 +519,14 @@ static void action_waitall(const char *const *action){ static void action_barrier(const char *const *action){ double clock = smpi_process_simulated_elapsed(); #ifdef HAVE_TRACING - int rank = smpi_comm_rank(MPI_COMM_WORLD); - TRACE_smpi_computing_out(rank); - TRACE_smpi_collective_in(rank, -1, __FUNCTION__, smpi_comm_size(MPI_COMM_WORLD)); + int rank = smpi_process_index(); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_BARRIER; + TRACE_smpi_collective_in(rank, -1, __FUNCTION__, extra); #endif - smpi_mpi_barrier(MPI_COMM_WORLD); + mpi_coll_barrier_fun(MPI_COMM_WORLD); #ifdef HAVE_TRACING TRACE_smpi_collective_out(rank, -1, __FUNCTION__); - TRACE_smpi_computing_in(rank); #endif log_timed_action (action, clock); @@ -394,6 +535,7 @@ static void action_barrier(const char *const *action){ static void action_bcast(const char *const *action) { + CHECK_ACTION_PARAMS(action, 1, 2); double size = parse_double(action[2]); double clock = smpi_process_simulated_elapsed(); int root=0; @@ -411,23 +553,28 @@ static void action_bcast(const char *const *action) } #ifdef HAVE_TRACING - int rank = smpi_comm_rank(MPI_COMM_WORLD); - TRACE_smpi_computing_out(rank); - int root_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), 0); - TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__,size*smpi_datatype_size(MPI_CURRENT_TYPE)); -#endif + int rank = smpi_process_index(); + int root_traced = smpi_group_index(smpi_comm_group(MPI_COMM_WORLD), root); - smpi_mpi_bcast(NULL, size, MPI_CURRENT_TYPE, root, MPI_COMM_WORLD); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_BCAST; + extra->send_size = size; + extra->root = root_traced; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__, extra); + +#endif + void *sendbuf = smpi_get_tmp_sendbuffer(size* smpi_datatype_size(MPI_CURRENT_TYPE)); + mpi_coll_bcast_fun(sendbuf, size, MPI_CURRENT_TYPE, root, MPI_COMM_WORLD); #ifdef HAVE_TRACING TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__); - TRACE_smpi_computing_in(rank); #endif - log_timed_action (action, clock); } static void action_reduce(const char *const *action) { + CHECK_ACTION_PARAMS(action, 2, 2); double comm_size = parse_double(action[2]); double comp_size = parse_double(action[3]); double clock = smpi_process_simulated_elapsed(); @@ -440,24 +587,33 @@ static void action_reduce(const char *const *action) MPI_CURRENT_TYPE=decode_datatype(action[5]); } } + + #ifdef HAVE_TRACING - int rank = smpi_comm_rank(MPI_COMM_WORLD); - TRACE_smpi_computing_out(rank); - int root_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), 0); - TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__,comm_size*smpi_datatype_size(MPI_CURRENT_TYPE)); + int rank = smpi_process_index(); + int root_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), root); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_REDUCE; + extra->send_size = comm_size; + extra->comp_size = comp_size; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + extra->root = root_traced; + + TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__,extra); #endif - mpi_coll_reduce_fun(NULL, NULL, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, root, MPI_COMM_WORLD); + void *recvbuf = smpi_get_tmp_sendbuffer(comm_size* smpi_datatype_size(MPI_CURRENT_TYPE)); + void *sendbuf = smpi_get_tmp_sendbuffer(comm_size* smpi_datatype_size(MPI_CURRENT_TYPE)); + mpi_coll_reduce_fun(sendbuf, recvbuf, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, root, MPI_COMM_WORLD); smpi_execute_flops(comp_size); #ifdef HAVE_TRACING TRACE_smpi_collective_out(rank, root_traced, __FUNCTION__); - TRACE_smpi_computing_in(rank); #endif - log_timed_action (action, clock); } static void action_allReduce(const char *const *action) { + CHECK_ACTION_PARAMS(action, 2, 1); double comm_size = parse_double(action[2]); double comp_size = parse_double(action[3]); @@ -466,18 +622,22 @@ static void action_allReduce(const char *const *action) { double clock = smpi_process_simulated_elapsed(); #ifdef HAVE_TRACING - int rank = smpi_comm_rank(MPI_COMM_WORLD); - TRACE_smpi_computing_out(rank); - TRACE_smpi_collective_in(rank, -1, __FUNCTION__,comp_size*smpi_datatype_size(MPI_CURRENT_TYPE)); + int rank = smpi_process_index(); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_ALLREDUCE; + extra->send_size = comm_size; + extra->comp_size = comp_size; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + + TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra); #endif - mpi_coll_reduce_fun(NULL, NULL, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, 0, MPI_COMM_WORLD); + void *recvbuf = smpi_get_tmp_sendbuffer(comm_size* smpi_datatype_size(MPI_CURRENT_TYPE)); + void *sendbuf = smpi_get_tmp_sendbuffer(comm_size* smpi_datatype_size(MPI_CURRENT_TYPE)); + mpi_coll_allreduce_fun(sendbuf, recvbuf, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, MPI_COMM_WORLD); smpi_execute_flops(comp_size); - mpi_coll_bcast_fun(NULL, comm_size, MPI_CURRENT_TYPE, 0, MPI_COMM_WORLD); #ifdef HAVE_TRACING TRACE_smpi_collective_out(rank, -1, __FUNCTION__); - TRACE_smpi_computing_in(rank); #endif - log_timed_action (action, clock); } @@ -496,25 +656,28 @@ static void action_allToAll(const char *const *action) { MPI_CURRENT_TYPE=MPI_DEFAULT_TYPE; MPI_CURRENT_TYPE2=MPI_DEFAULT_TYPE; } - void *send = calloc(send_size*comm_size, smpi_datatype_size(MPI_CURRENT_TYPE)); - void *recv = calloc(recv_size*comm_size, smpi_datatype_size(MPI_CURRENT_TYPE2)); + void *send = smpi_get_tmp_sendbuffer(send_size*comm_size* smpi_datatype_size(MPI_CURRENT_TYPE)); + void *recv = smpi_get_tmp_recvbuffer(recv_size*comm_size* smpi_datatype_size(MPI_CURRENT_TYPE2)); #ifdef HAVE_TRACING int rank = smpi_process_index(); - TRACE_smpi_computing_out(rank); - TRACE_smpi_collective_in(rank, -1, __FUNCTION__,send_size*smpi_datatype_size(MPI_CURRENT_TYPE)); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_ALLTOALL; + extra->send_size = send_size; + extra->recv_size = recv_size; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + extra->datatype2 = encode_datatype(MPI_CURRENT_TYPE2, NULL); + + TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra); #endif mpi_coll_alltoall_fun(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, MPI_COMM_WORLD); #ifdef HAVE_TRACING TRACE_smpi_collective_out(rank, -1, __FUNCTION__); - TRACE_smpi_computing_in(rank); #endif - log_timed_action (action, clock); - xbt_free(send); - xbt_free(recv); + } @@ -531,6 +694,7 @@ static void action_gather(const char *const *action) { 4) 0 is the send datatype id, see decode_datatype() 5) 0 is the recv datatype id, see decode_datatype() */ + CHECK_ACTION_PARAMS(action, 2, 3); double clock = smpi_process_simulated_elapsed(); int comm_size = smpi_comm_size(MPI_COMM_WORLD); int send_size = parse_double(action[2]); @@ -543,34 +707,111 @@ static void action_gather(const char *const *action) { MPI_CURRENT_TYPE=MPI_DEFAULT_TYPE; MPI_CURRENT_TYPE2=MPI_DEFAULT_TYPE; } - void *send = calloc(send_size, smpi_datatype_size(MPI_CURRENT_TYPE)); - void *recv = calloc(recv_size, smpi_datatype_size(MPI_CURRENT_TYPE2)); - - int root=atoi(action[4]); - int rank = smpi_process_index(); + void *send = smpi_get_tmp_sendbuffer(send_size* smpi_datatype_size(MPI_CURRENT_TYPE)); + void *recv = NULL; + int root=0; + if(action[4]) + root=atoi(action[4]); + int rank = smpi_comm_rank(MPI_COMM_WORLD); if(rank==root) - recv = calloc(recv_size*comm_size, smpi_datatype_size(MPI_CURRENT_TYPE2)); + recv = smpi_get_tmp_recvbuffer(recv_size*comm_size* smpi_datatype_size(MPI_CURRENT_TYPE2)); #ifdef HAVE_TRACING - TRACE_smpi_computing_out(rank); - TRACE_smpi_collective_in(rank, -1, __FUNCTION__,send_size*smpi_datatype_size(MPI_CURRENT_TYPE)); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_GATHER; + extra->send_size = send_size; + extra->recv_size = recv_size; + extra->root = root; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + extra->datatype2 = encode_datatype(MPI_CURRENT_TYPE2, NULL); + + TRACE_smpi_collective_in(smpi_process_index(), root, __FUNCTION__, extra); #endif -smpi_mpi_gather(send, send_size, MPI_CURRENT_TYPE, + mpi_coll_gather_fun(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, root, MPI_COMM_WORLD); #ifdef HAVE_TRACING - TRACE_smpi_collective_out(rank, -1, __FUNCTION__); - TRACE_smpi_computing_in(rank); + TRACE_smpi_collective_out(smpi_process_index(), -1, __FUNCTION__); #endif - log_timed_action (action, clock); - xbt_free(send); - xbt_free(recv); + } + +static void action_gatherv(const char *const *action) { + /* + The structure of the gatherv action for the rank 0 (total 4 processes) + is the following: + 0 gather 68 68 10 10 10 0 0 0 + + where: + 1) 68 is the sendcount + 2) 68 10 10 10 is the recvcounts + 3) 0 is the root node + 4) 0 is the send datatype id, see decode_datatype() + 5) 0 is the recv datatype id, see decode_datatype() + */ + + double clock = smpi_process_simulated_elapsed(); + int comm_size = smpi_comm_size(MPI_COMM_WORLD); + CHECK_ACTION_PARAMS(action, comm_size+1, 2); + int send_size = parse_double(action[2]); + int *disps = xbt_new0(int, comm_size); + int *recvcounts = xbt_new0(int, comm_size); + int i=0,recv_sum=0; + + MPI_Datatype MPI_CURRENT_TYPE2; + if(action[4+comm_size]) { + MPI_CURRENT_TYPE=decode_datatype(action[4+comm_size]); + MPI_CURRENT_TYPE2=decode_datatype(action[5+comm_size]); + } else { + MPI_CURRENT_TYPE=MPI_DEFAULT_TYPE; + MPI_CURRENT_TYPE2=MPI_DEFAULT_TYPE; + } + void *send = smpi_get_tmp_sendbuffer(send_size* smpi_datatype_size(MPI_CURRENT_TYPE)); + void *recv = NULL; + for(i=0;itype = TRACING_GATHERV; + extra->send_size = send_size; + extra->recvcounts= xbt_malloc(comm_size*sizeof(int)); + for(i=0; i< comm_size; i++)//copy data to avoid bad free + extra->recvcounts[i] = recvcounts[i]; + extra->root = root; + extra->num_processes = comm_size; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + extra->datatype2 = encode_datatype(MPI_CURRENT_TYPE2, NULL); + + TRACE_smpi_collective_in(smpi_process_index(), root, __FUNCTION__, extra); +#endif +smpi_mpi_gatherv(send, send_size, MPI_CURRENT_TYPE, + recv, recvcounts, disps, MPI_CURRENT_TYPE2, + root, MPI_COMM_WORLD); + +#ifdef HAVE_TRACING + TRACE_smpi_collective_out(smpi_process_index(), -1, __FUNCTION__); +#endif + + log_timed_action (action, clock); + xbt_free(recvcounts); + xbt_free(disps); +} + static void action_reducescatter(const char *const *action) { /* @@ -589,13 +830,13 @@ static void action_reducescatter(const char *const *action) { double clock = smpi_process_simulated_elapsed(); int comm_size = smpi_comm_size(MPI_COMM_WORLD); + CHECK_ACTION_PARAMS(action, comm_size+1, 1); int comp_size = parse_double(action[2+comm_size]); int *recvcounts = xbt_new0(int, comm_size); int *disps = xbt_new0(int, comm_size); - int i=0,recv_sum=0; - int root=0; + int i=0; int rank = smpi_process_index(); - + int size = 0; if(action[3+comm_size]) MPI_CURRENT_TYPE=decode_datatype(action[3+comm_size]); else @@ -603,29 +844,92 @@ static void action_reducescatter(const char *const *action) { for(i=0;itype = TRACING_REDUCE_SCATTER; + extra->send_size = 0; + extra->recvcounts= xbt_malloc(comm_size*sizeof(int)); + for(i=0; i< comm_size; i++)//copy data to avoid bad free + extra->recvcounts[i] = recvcounts[i]; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + extra->comp_size = comp_size; + extra->num_processes = comm_size; + + TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra); #endif - mpi_coll_reduce_fun(NULL, NULL, recv_sum, MPI_CURRENT_TYPE, MPI_OP_NULL, - root, MPI_COMM_WORLD); - smpi_mpi_scatterv(NULL, recvcounts, disps, MPI_CURRENT_TYPE, NULL, - recvcounts[rank], MPI_CURRENT_TYPE, 0, MPI_COMM_WORLD); + void *sendbuf = smpi_get_tmp_sendbuffer(size* smpi_datatype_size(MPI_CURRENT_TYPE)); + void *recvbuf = smpi_get_tmp_recvbuffer(size* smpi_datatype_size(MPI_CURRENT_TYPE)); + + mpi_coll_reduce_scatter_fun(sendbuf, recvbuf, recvcounts, MPI_CURRENT_TYPE, MPI_OP_NULL, + MPI_COMM_WORLD); smpi_execute_flops(comp_size); #ifdef HAVE_TRACING TRACE_smpi_collective_out(rank, -1, __FUNCTION__); - TRACE_smpi_computing_in(rank); #endif - + xbt_free(recvcounts); + xbt_free(disps); log_timed_action (action, clock); } +static void action_allgather(const char *const *action) { + /* + The structure of the allgather action for the rank 0 (total 4 processes) + is the following: + 0 allGather 275427 275427 + + where: + 1) 275427 is the sendcount + 2) 275427 is the recvcount + 3) No more values mean that the datatype for sent and receive buffer + is the default one, see decode_datatype(). + + */ + + double clock = smpi_process_simulated_elapsed(); + + CHECK_ACTION_PARAMS(action, 2, 2); + int sendcount=atoi(action[2]); + int recvcount=atoi(action[3]); + + MPI_Datatype MPI_CURRENT_TYPE2; + + if(action[4]) { + MPI_CURRENT_TYPE = decode_datatype(action[3]); + MPI_CURRENT_TYPE2 = decode_datatype(action[4]); + } else { + MPI_CURRENT_TYPE = MPI_DEFAULT_TYPE; + MPI_CURRENT_TYPE2 = MPI_DEFAULT_TYPE; + } + void *sendbuf = smpi_get_tmp_sendbuffer(sendcount* smpi_datatype_size(MPI_CURRENT_TYPE)); + void *recvbuf = smpi_get_tmp_recvbuffer(recvcount* smpi_datatype_size(MPI_CURRENT_TYPE2)); + +#ifdef HAVE_TRACING + int rank = smpi_process_index(); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_ALLGATHER; + extra->send_size = sendcount; + extra->recv_size= recvcount; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + extra->datatype2 = encode_datatype(MPI_CURRENT_TYPE2, NULL); + extra->num_processes = smpi_comm_size(MPI_COMM_WORLD); + + TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra); +#endif + + mpi_coll_allgather_fun(sendbuf, sendcount, MPI_CURRENT_TYPE, recvbuf, recvcount, MPI_CURRENT_TYPE2, MPI_COMM_WORLD); + +#ifdef HAVE_TRACING + TRACE_smpi_collective_out(rank, -1, __FUNCTION__); +#endif + + log_timed_action (action, clock); +} static void action_allgatherv(const char *const *action) { @@ -645,6 +949,7 @@ static void action_allgatherv(const char *const *action) { double clock = smpi_process_simulated_elapsed(); int comm_size = smpi_comm_size(MPI_COMM_WORLD); + CHECK_ACTION_PARAMS(action, comm_size+1, 2); int i=0; int sendcount=atoi(action[2]); int *recvcounts = xbt_new0(int, comm_size); @@ -659,35 +964,40 @@ static void action_allgatherv(const char *const *action) { MPI_CURRENT_TYPE = MPI_DEFAULT_TYPE; MPI_CURRENT_TYPE2 = MPI_DEFAULT_TYPE; } - void *sendbuf = calloc(sendcount, smpi_datatype_size(MPI_CURRENT_TYPE)); + void *sendbuf = smpi_get_tmp_sendbuffer(sendcount* smpi_datatype_size(MPI_CURRENT_TYPE)); for(i=0;itype = TRACING_ALLGATHERV; + extra->send_size = sendcount; + extra->recvcounts= xbt_malloc(comm_size*sizeof(int)); + for(i=0; i< comm_size; i++)//copy data to avoid bad free + extra->recvcounts[i] = recvcounts[i]; + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + extra->datatype2 = encode_datatype(MPI_CURRENT_TYPE2, NULL); + extra->num_processes = comm_size; + + TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra); #endif -mpi_coll_allgatherv_fun(sendbuf, sendcount, MPI_CURRENT_TYPE, recvbuf, recvcounts, disps, MPI_CURRENT_TYPE2, MPI_COMM_WORLD); + mpi_coll_allgatherv_fun(sendbuf, sendcount, MPI_CURRENT_TYPE, recvbuf, recvcounts, disps, MPI_CURRENT_TYPE2, MPI_COMM_WORLD); #ifdef HAVE_TRACING TRACE_smpi_collective_out(rank, -1, __FUNCTION__); - TRACE_smpi_computing_in(rank); #endif log_timed_action (action, clock); - xbt_free(sendbuf); - xbt_free(recvbuf); xbt_free(recvcounts); xbt_free(disps); } - static void action_allToAllv(const char *const *action) { /* The structure of the allToAllV action for the rank 0 (total 4 processes) @@ -706,6 +1016,7 @@ static void action_allToAllv(const char *const *action) { double clock = smpi_process_simulated_elapsed(); int comm_size = smpi_comm_size(MPI_COMM_WORLD); + CHECK_ACTION_PARAMS(action, 2*comm_size+2, 2); int send_buf_size=0,recv_buf_size=0,i=0; int *sendcounts = xbt_new0(int, comm_size); int *recvcounts = xbt_new0(int, comm_size); @@ -725,8 +1036,8 @@ static void action_allToAllv(const char *const *action) { MPI_CURRENT_TYPE2=MPI_DEFAULT_TYPE; } - void *sendbuf = calloc(send_buf_size, smpi_datatype_size(MPI_CURRENT_TYPE)); - void *recvbuf = calloc(recv_buf_size, smpi_datatype_size(MPI_CURRENT_TYPE2)); + void *sendbuf = smpi_get_tmp_sendbuffer(send_buf_size* smpi_datatype_size(MPI_CURRENT_TYPE)); + void *recvbuf = smpi_get_tmp_recvbuffer(recv_buf_size* smpi_datatype_size(MPI_CURRENT_TYPE2)); for(i=0;itype = TRACING_ALLTOALLV; + extra->recvcounts= xbt_malloc(comm_size*sizeof(int)); + extra->sendcounts= xbt_malloc(comm_size*sizeof(int)); + extra->num_processes = comm_size; + + for(i=0; i< comm_size; i++){//copy data to avoid bad free + extra->send_size += sendcounts[i]; + extra->sendcounts[i] = sendcounts[i]; + extra->recv_size += recvcounts[i]; + extra->recvcounts[i] = recvcounts[i]; + } + extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE, NULL); + extra->datatype2 = encode_datatype(MPI_CURRENT_TYPE2, NULL); + + TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra); #endif - mpi_coll_alltoallv_fun(sendbuf, sendcounts, senddisps, MPI_CURRENT_TYPE, + mpi_coll_alltoallv_fun(sendbuf, sendcounts, senddisps, MPI_CURRENT_TYPE, recvbuf, recvcounts, recvdisps, MPI_CURRENT_TYPE, MPI_COMM_WORLD); #ifdef HAVE_TRACING TRACE_smpi_collective_out(rank, -1, __FUNCTION__); - TRACE_smpi_computing_in(rank); #endif log_timed_action (action, clock); - xbt_free(sendbuf); - xbt_free(recvbuf); xbt_free(sendcounts); xbt_free(recvcounts); xbt_free(senddisps); @@ -759,7 +1079,19 @@ static void action_allToAllv(const char *const *action) { } void smpi_replay_init(int *argc, char***argv){ - PMPI_Init(argc, argv); + smpi_process_init(argc, argv); + smpi_process_mark_as_initialized(); + smpi_process_set_replaying(1); +#ifdef HAVE_TRACING + int rank = smpi_process_index(); + TRACE_smpi_init(rank); + TRACE_smpi_computing_init(rank); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_INIT; + TRACE_smpi_collective_in(rank, -1, __FUNCTION__, extra); + TRACE_smpi_collective_out(rank, -1, __FUNCTION__); +#endif + if (!smpi_process_index()){ _xbt_replay_action_init(); xbt_replay_action_register("init", action_init); @@ -771,6 +1103,7 @@ void smpi_replay_init(int *argc, char***argv){ xbt_replay_action_register("Isend", action_Isend); xbt_replay_action_register("recv", action_recv); xbt_replay_action_register("Irecv", action_Irecv); + xbt_replay_action_register("test", action_test); xbt_replay_action_register("wait", action_wait); xbt_replay_action_register("waitAll", action_waitall); xbt_replay_action_register("barrier", action_barrier); @@ -780,30 +1113,73 @@ void smpi_replay_init(int *argc, char***argv){ xbt_replay_action_register("allToAll", action_allToAll); xbt_replay_action_register("allToAllV", action_allToAllv); xbt_replay_action_register("gather", action_gather); + xbt_replay_action_register("gatherV", action_gatherv); + xbt_replay_action_register("allGather", action_allgather); xbt_replay_action_register("allGatherV", action_allgatherv); xbt_replay_action_register("reduceScatter", action_reducescatter); xbt_replay_action_register("compute", action_compute); } - + + //if we have a delayed start, sleep here. + if(*argc>2){ + char *endptr; + double value = strtod((*argv)[2], &endptr); + if (*endptr != '\0') + THROWF(unknown_error, 0, "%s is not a double", (*argv)[2]); + XBT_VERB("Delayed start for instance - Sleeping for %f flops ",value ); + smpi_execute_flops(value); + } xbt_replay_action_runner(*argc, *argv); } int smpi_replay_finalize(){ double sim_time= 1.; /* One active process will stop. Decrease the counter*/ - active_processes--; XBT_DEBUG("There are %lu elements in reqq[*]", - xbt_dynar_length(reqq[smpi_comm_rank(MPI_COMM_WORLD)])); - xbt_dynar_free(&reqq[smpi_comm_rank(MPI_COMM_WORLD)]); + xbt_dynar_length(reqq[smpi_process_index()])); + if (!xbt_dynar_is_empty(reqq[smpi_process_index()])){ + int count_requests=xbt_dynar_length(reqq[smpi_process_index()]); + MPI_Request requests[count_requests]; + MPI_Status status[count_requests]; + unsigned int i; + + xbt_dynar_foreach(reqq[smpi_process_index()],i,requests[i]); + smpi_mpi_waitall(count_requests, requests, status); + active_processes--; + } else { + active_processes--; + } + if(!active_processes){ /* Last process alive speaking */ /* end the simulated timer */ sim_time = smpi_process_simulated_elapsed(); - XBT_INFO("Simulation time %g", sim_time); + } + + + xbt_dynar_free_container(&(reqq[smpi_process_index()])); + + if(!active_processes){ + XBT_INFO("Simulation time %f", sim_time); _xbt_replay_action_exit(); + xbt_free(sendbuffer); + xbt_free(recvbuffer); xbt_free(reqq); reqq = NULL; } - smpi_mpi_barrier(MPI_COMM_WORLD); - return PMPI_Finalize(); + + +#ifdef HAVE_TRACING + int rank = smpi_process_index(); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type = TRACING_FINALIZE; + TRACE_smpi_collective_in(rank, -1, __FUNCTION__, extra); +#endif + smpi_process_finalize(); +#ifdef HAVE_TRACING + TRACE_smpi_collective_out(rank, -1, __FUNCTION__); + TRACE_smpi_finalize(smpi_process_index()); +#endif + smpi_process_destroy(); + return MPI_SUCCESS; }