X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/8a7d5cba17741cb4a0c0b66ac5ab52871f100d6a..cd69f8c66886ae218c6a9b38125b27fe157e6bb2:/src/smpi/smpi_replay.c diff --git a/src/smpi/smpi_replay.c b/src/smpi/smpi_replay.c index e8f0ada9ad..09c5b6be1c 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 */ @@ -125,21 +158,13 @@ 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) @@ -166,10 +191,11 @@ 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(); double flops= parse_double(action[2]); #ifdef HAVE_TRACING - int rank = smpi_comm_rank(MPI_COMM_WORLD); + 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; @@ -185,6 +211,7 @@ static void action_compute(const char *const *action) 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(); @@ -196,7 +223,7 @@ static void action_send(const char *const *action) } #ifdef HAVE_TRACING - int rank = smpi_comm_rank(MPI_COMM_WORLD); + int rank = smpi_process_index(); int dst_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), to); instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); @@ -221,6 +248,7 @@ static void action_send(const char *const *action) 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(); @@ -230,7 +258,7 @@ 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); + int rank = smpi_process_index(); int dst_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), to); instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); extra->type = TRACING_ISEND; @@ -249,12 +277,13 @@ static void action_Isend(const char *const *action) request->send = 1; #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(); @@ -264,7 +293,7 @@ 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); instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); @@ -276,6 +305,12 @@ static void action_recv(const char *const *action) { 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 @@ -288,19 +323,17 @@ 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); instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); extra->type = TRACING_IRECV; @@ -310,6 +343,12 @@ static void action_Irecv(const char *const *action) extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE); 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); @@ -317,25 +356,65 @@ 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; @@ -359,11 +438,12 @@ static void action_wait(const char *const *action){ } 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]; @@ -371,7 +451,7 @@ 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 @@ -427,7 +507,9 @@ static void action_waitall(const char *const *action){ xbt_dynar_free(&recvs); #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); } @@ -435,12 +517,12 @@ 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); + 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__); #endif @@ -451,6 +533,7 @@ 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; @@ -468,7 +551,7 @@ static void action_bcast(const char *const *action) } #ifdef HAVE_TRACING - int rank = smpi_comm_rank(MPI_COMM_WORLD); + int rank = smpi_process_index(); int root_traced = smpi_group_index(smpi_comm_group(MPI_COMM_WORLD), root); instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); @@ -479,17 +562,17 @@ static void action_bcast(const char *const *action) TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__, extra); #endif - - mpi_coll_bcast_fun(NULL, size, MPI_CURRENT_TYPE, root, MPI_COMM_WORLD); + 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__); #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(); @@ -502,9 +585,11 @@ 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); + 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; @@ -515,16 +600,18 @@ static void action_reduce(const char *const *action) 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__); #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]); @@ -533,7 +620,7 @@ 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); + 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; @@ -542,13 +629,13 @@ static void action_allReduce(const char *const *action) { 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__); #endif - log_timed_action (action, clock); } @@ -567,8 +654,8 @@ 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(); @@ -587,10 +674,8 @@ static void action_allToAll(const char *const *action) { #ifdef HAVE_TRACING TRACE_smpi_collective_out(rank, -1, __FUNCTION__); #endif - log_timed_action (action, clock); - xbt_free(send); - xbt_free(recv); + } @@ -607,6 +692,7 @@ 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]); @@ -619,14 +705,15 @@ 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 *send = smpi_get_tmp_sendbuffer(send_size* smpi_datatype_size(MPI_CURRENT_TYPE)); void *recv = NULL; - - int root=atoi(action[4]); - int rank = smpi_process_index(); + 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 instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); @@ -637,19 +724,17 @@ static void action_gather(const char *const *action) { extra->datatype1 = encode_datatype(MPI_CURRENT_TYPE); extra->datatype2 = encode_datatype(MPI_CURRENT_TYPE2); - TRACE_smpi_collective_in(rank, root, __FUNCTION__, extra); + 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_collective_out(smpi_process_index(), -1, __FUNCTION__); #endif - log_timed_action (action, clock); - xbt_free(send); - xbt_free(recv); + } @@ -667,8 +752,10 @@ static void action_gatherv(const char *const *action) { 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); @@ -682,7 +769,7 @@ static void action_gatherv(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 *send = smpi_get_tmp_sendbuffer(send_size* smpi_datatype_size(MPI_CURRENT_TYPE)); void *recv = NULL; for(i=0;idatatype1 = encode_datatype(MPI_CURRENT_TYPE); extra->datatype2 = encode_datatype(MPI_CURRENT_TYPE2); - TRACE_smpi_collective_in(rank, root, __FUNCTION__, extra); + 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(rank, -1, __FUNCTION__); + TRACE_smpi_collective_out(smpi_process_index(), -1, __FUNCTION__); #endif log_timed_action (action, clock); xbt_free(recvcounts); - xbt_free(send); - if(recv)xbt_free(recv); xbt_free(disps); - } static void action_reducescatter(const char *const *action) { @@ -744,13 +828,13 @@ 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 @@ -758,8 +842,8 @@ static void action_reducescatter(const char *const *action) { for(i=0;icomp_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); @@ -791,6 +875,59 @@ static void action_reducescatter(const char *const *action) { 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); + extra->datatype2 = encode_datatype(MPI_CURRENT_TYPE2); + 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) { @@ -810,6 +947,7 @@ 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); @@ -824,13 +962,13 @@ 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;i2){ + 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); + + #ifdef HAVE_TRACING int rank = smpi_process_index(); instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);