X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/b05a0ed4f5fb7b81f96777af37b8f7fc5a4bf160..a9b0436a2cbb60596670309f9e80229bccbfd05c:/src/smpi/smpi_replay.c diff --git a/src/smpi/smpi_replay.c b/src/smpi/smpi_replay.c index dc35a680d3..d7990bf2ab 100644 --- a/src/smpi/smpi_replay.c +++ b/src/smpi/smpi_replay.c @@ -13,12 +13,25 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_replay,smpi,"Trace Replay with SMPI"); int communicator_size = 0; +static int active_processes = 0; +xbt_dynar_t *reqq; + +MPI_Datatype MPI_DEFAULT_TYPE, MPI_CURRENT_TYPE; + +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, " "); + XBT_VERB("%s %f", name, smpi_process_simulated_elapsed()-clock); + free(name); + } +} typedef struct { xbt_dynar_t isends; /* of MPI_Request */ xbt_dynar_t irecvs; /* of MPI_Request */ } s_smpi_replay_globals_t, *smpi_replay_globals_t; + /* Helper function */ static double parse_double(const char *string) { @@ -30,25 +43,69 @@ static double parse_double(const char *string) return value; } +static MPI_Datatype decode_datatype(const char *const action) +{ +// Declared datatypes, + + switch(atoi(action)) + { + case 0: + MPI_CURRENT_TYPE=MPI_DOUBLE; + break; + case 1: + MPI_CURRENT_TYPE=MPI_INT; + break; + case 2: + MPI_CURRENT_TYPE=MPI_CHAR; + break; + case 3: + MPI_CURRENT_TYPE=MPI_SHORT; + break; + case 4: + MPI_CURRENT_TYPE=MPI_LONG; + break; + case 5: + MPI_CURRENT_TYPE=MPI_FLOAT; + break; + case 6: + MPI_CURRENT_TYPE=MPI_BYTE; + break; + default: + MPI_CURRENT_TYPE=MPI_DEFAULT_TYPE; + + } + return MPI_CURRENT_TYPE; +} + static void action_init(const char *const *action) { + int i; XBT_DEBUG("Initialize the counters"); smpi_replay_globals_t globals = xbt_new(s_smpi_replay_globals_t, 1); globals->isends = xbt_dynar_new(sizeof(MPI_Request),NULL); globals->irecvs = xbt_dynar_new(sizeof(MPI_Request),NULL); + 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 */ + active_processes = smpi_process_count(); + + reqq=xbt_new0(xbt_dynar_t,active_processes); + + for(i=0;iisends),xbt_dynar_length(globals->irecvs)); @@ -56,12 +113,6 @@ static void action_finalize(const char *const *action) xbt_dynar_free_container(&(globals->irecvs)); } free(globals); - /* end the simulated timer */ - sim_time = smpi_process_simulated_elapsed(); - if (!smpi_process_index()) - XBT_INFO("Simulation time %g", sim_time); - smpi_process_finalize(); - smpi_process_destroy(); } static void action_comm_size(const char *const *action) @@ -69,19 +120,29 @@ static void action_comm_size(const char *const *action) double clock = smpi_process_simulated_elapsed(); communicator_size = parse_double(action[2]); + log_timed_action (action, clock); +} - XBT_VERB("%s %f", xbt_str_join_array(action, " "), - smpi_process_simulated_elapsed()-clock); +static void action_comm_split(const char *const *action) +{ + double clock = smpi_process_simulated_elapsed(); + + log_timed_action (action, clock); } +static void action_comm_dup(const char *const *action) +{ + double clock = smpi_process_simulated_elapsed(); + + log_timed_action (action, clock); +} static void action_compute(const char *const *action) { double clock = smpi_process_simulated_elapsed(); smpi_execute_flops(parse_double(action[2])); - XBT_VERB("%s %f", xbt_str_join_array(action, " "), - smpi_process_simulated_elapsed()-clock); + log_timed_action (action, clock); } static void action_send(const char *const *action) @@ -89,6 +150,13 @@ static void action_send(const char *const *action) int to = atoi(action[2]); double size=parse_double(action[3]); double clock = smpi_process_simulated_elapsed(); + + 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); TRACE_smpi_computing_out(rank); @@ -97,10 +165,11 @@ static void action_send(const char *const *action) TRACE_smpi_send(rank, rank, dst_traced); #endif - smpi_mpi_send(NULL, size, MPI_BYTE, to , 0, MPI_COMM_WORLD); - XBT_VERB("%s %f", xbt_str_join_array(action, " "), - smpi_process_simulated_elapsed()-clock); -#ifdef HAVE_TRACING + smpi_mpi_send(NULL, size, MPI_CURRENT_TYPE, to , 0, MPI_COMM_WORLD); + + log_timed_action (action, clock); + + #ifdef HAVE_TRACING TRACE_smpi_ptp_out(rank, rank, dst_traced, __FUNCTION__); TRACE_smpi_computing_in(rank); #endif @@ -112,6 +181,11 @@ static void action_Isend(const char *const *action) int to = atoi(action[2]); double size=parse_double(action[3]); double clock = smpi_process_simulated_elapsed(); + MPI_Request request; + + if(action[4]) MPI_CURRENT_TYPE=decode_datatype(action[4]); + else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; + smpi_replay_globals_t globals = (smpi_replay_globals_t) smpi_process_get_user_data(); #ifdef HAVE_TRACING @@ -122,8 +196,8 @@ static void action_Isend(const char *const *action) TRACE_smpi_send(rank, rank, dst_traced); #endif - MPI_Request request = smpi_mpi_isend(NULL, size, MPI_BYTE, to, 0, - MPI_COMM_WORLD); + request = smpi_mpi_isend(NULL, size, MPI_CURRENT_TYPE, to, 0,MPI_COMM_WORLD); + #ifdef HAVE_TRACING TRACE_smpi_ptp_out(rank, rank, dst_traced, __FUNCTION__); request->send = 1; @@ -131,10 +205,9 @@ static void action_Isend(const char *const *action) #endif xbt_dynar_push(globals->isends,&request); + xbt_dynar_push(reqq[smpi_comm_rank(MPI_COMM_WORLD)],&request); - //TODO do the asynchronous cleanup - XBT_VERB("%s %f", xbt_str_join_array(action, " "), - smpi_process_simulated_elapsed()-clock); + log_timed_action (action, clock); } static void action_recv(const char *const *action) { @@ -142,6 +215,10 @@ static void action_recv(const char *const *action) { double size=parse_double(action[3]); double clock = smpi_process_simulated_elapsed(); MPI_Status status; + + 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 src_traced = smpi_group_rank(smpi_comm_group(MPI_COMM_WORLD), from); @@ -150,7 +227,7 @@ static void action_recv(const char *const *action) { TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__); #endif - smpi_mpi_recv(NULL, size, MPI_BYTE, from, 0, MPI_COMM_WORLD, &status); + 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__); @@ -158,8 +235,7 @@ static void action_recv(const char *const *action) { TRACE_smpi_computing_in(rank); #endif - XBT_VERB("%s %f", xbt_str_join_array(action, " "), - smpi_process_simulated_elapsed()-clock); + log_timed_action (action, clock); } static void action_Irecv(const char *const *action) @@ -168,8 +244,12 @@ static void action_Irecv(const char *const *action) 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); @@ -177,16 +257,16 @@ static void action_Irecv(const char *const *action) TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__); #endif - request = smpi_mpi_irecv(NULL, size, MPI_BYTE, from, 0, MPI_COMM_WORLD); + request = smpi_mpi_irecv(NULL, size, MPI_CURRENT_TYPE, from, 0, MPI_COMM_WORLD); + #ifdef HAVE_TRACING 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); - //TODO do the asynchronous cleanup - XBT_VERB("%s %f", xbt_str_join_array(action, " "), - smpi_process_simulated_elapsed()-clock); + log_timed_action (action, clock); } static void action_wait(const char *const *action){ @@ -221,8 +301,81 @@ static void action_wait(const char *const *action){ TRACE_smpi_computing_in(rank); #endif - XBT_VERB("%s %f", xbt_str_join_array(action, " "), - smpi_process_simulated_elapsed()-clock); + log_timed_action (action, clock); +} + +static void action_waitall(const char *const *action){ + 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)]); + + if (count_requests>0) { + MPI_Request requests[count_requests]; + MPI_Status status[count_requests]; + + /* 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]); + + #ifdef HAVE_TRACING + //save information from requests + + xbt_dynar_t srcs = xbt_dynar_new(sizeof(int), NULL); + xbt_dynar_t dsts = xbt_dynar_new(sizeof(int), NULL); + xbt_dynar_t recvs = xbt_dynar_new(sizeof(int), NULL); + for (i = 0; i < count_requests; i++) { + if(requests[i]){ + int *asrc = xbt_new(int, 1); + int *adst = xbt_new(int, 1); + int *arecv = xbt_new(int, 1); + *asrc = requests[i]->src; + *adst = requests[i]->dst; + *arecv = requests[i]->recv; + xbt_dynar_insert_at(srcs, i, asrc); + xbt_dynar_insert_at(dsts, i, adst); + xbt_dynar_insert_at(recvs, i, arecv); + xbt_free(asrc); + xbt_free(adst); + xbt_free(arecv); + }else { + int *t = xbt_new(int, 1); + xbt_dynar_insert_at(srcs, i, t); + xbt_dynar_insert_at(dsts, i, t); + xbt_dynar_insert_at(recvs, i, t); + xbt_free(t); + } + } + int rank_traced = smpi_process_index(); + TRACE_smpi_computing_out(rank_traced); + + TRACE_smpi_ptp_in(rank_traced, -1, -1, __FUNCTION__); + #endif + + smpi_mpi_waitall(count_requests, requests, status); + + #ifdef HAVE_TRACING + for (i = 0; i < count_requests; i++) { + int src_traced, dst_traced, is_wait_for_receive; + xbt_dynar_get_cpy(srcs, i, &src_traced); + xbt_dynar_get_cpy(dsts, i, &dst_traced); + xbt_dynar_get_cpy(recvs, i, &is_wait_for_receive); + if (is_wait_for_receive) { + TRACE_smpi_recv(rank_traced, src_traced, dst_traced); + } + } + TRACE_smpi_ptp_out(rank_traced, -1, -1, __FUNCTION__); + //clean-up of dynars + 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)]); + } + log_timed_action (action, clock); } static void action_barrier(const char *const *action){ @@ -238,14 +391,28 @@ static void action_barrier(const char *const *action){ TRACE_smpi_computing_in(rank); #endif - XBT_VERB("%s %f", xbt_str_join_array(action, " "), - smpi_process_simulated_elapsed()-clock); + log_timed_action (action, clock); } + static void action_bcast(const char *const *action) { double size = parse_double(action[2]); double clock = smpi_process_simulated_elapsed(); + int root=0; + /* + * Initialize MPI_CURRENT_TYPE in order to decrease + * the number of the checks + * */ + MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; + + if(action[3]) { + root= atoi(action[3]); + if(action[4]) { + MPI_CURRENT_TYPE=decode_datatype(action[4]); + } + } + #ifdef HAVE_TRACING int rank = smpi_comm_rank(MPI_COMM_WORLD); TRACE_smpi_computing_out(rank); @@ -253,77 +420,216 @@ static void action_bcast(const char *const *action) TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__); #endif - smpi_mpi_bcast(NULL, size, MPI_BYTE, 0, MPI_COMM_WORLD); + smpi_mpi_bcast(NULL, 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 - XBT_VERB("%s %f", xbt_str_join_array(action, " "), - smpi_process_simulated_elapsed()-clock); + log_timed_action (action, clock); } static void action_reduce(const char *const *action) { - double size = parse_double(action[2]); + double comm_size = parse_double(action[2]); + double comp_size = parse_double(action[3]); double clock = smpi_process_simulated_elapsed(); + int root=0; + MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; + + if(action[4]) { + root= atoi(action[4]); + if(action[5]) { + 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__); #endif - smpi_mpi_reduce(NULL, NULL, size, MPI_BYTE, MPI_OP_NULL, 0, MPI_COMM_WORLD); + smpi_mpi_reduce(NULL, NULL, 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 - XBT_VERB("%s %f", xbt_str_join_array(action, " "), - smpi_process_simulated_elapsed()-clock); + log_timed_action (action, clock); } static void action_allReduce(const char *const *action) { double comm_size = parse_double(action[2]); double comp_size = parse_double(action[3]); + + if(action[4]) MPI_CURRENT_TYPE=decode_datatype(action[4]); + else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; + 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__); #endif - smpi_mpi_reduce(NULL, NULL, comm_size, MPI_BYTE, MPI_OP_NULL, 0, MPI_COMM_WORLD); + smpi_mpi_reduce(NULL, NULL, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, 0, MPI_COMM_WORLD); smpi_execute_flops(comp_size); - smpi_mpi_bcast(NULL, comm_size, MPI_BYTE, 0, MPI_COMM_WORLD); + smpi_mpi_bcast(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); +} + +static void action_allToAll(const char *const *action) { + double clock = smpi_process_simulated_elapsed(); + int comm_size = smpi_comm_size(MPI_COMM_WORLD); + int send_size = parse_double(action[2]); + int recv_size = parse_double(action[3]); + void *send = xbt_new0(int, send_size*comm_size); + void *recv = xbt_new0(int, send_size*comm_size); + + if(action[4]) MPI_CURRENT_TYPE=decode_datatype(action[4]); + else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; + +#ifdef HAVE_TRACING + int rank = smpi_process_index(); + TRACE_smpi_computing_out(rank); + TRACE_smpi_collective_in(rank, -1, __FUNCTION__); +#endif + + if (send_size < 200 && comm_size > 12) { + smpi_coll_tuned_alltoall_bruck(send, send_size, MPI_CURRENT_TYPE, + recv, recv_size, MPI_CURRENT_TYPE, + MPI_COMM_WORLD); + } else if (send_size < 3000) { + smpi_coll_tuned_alltoall_basic_linear(send, send_size, MPI_CURRENT_TYPE, + recv, recv_size, MPI_CURRENT_TYPE, + MPI_COMM_WORLD); + } else { + smpi_coll_tuned_alltoall_pairwise(send, send_size, MPI_CURRENT_TYPE, + recv, recv_size, MPI_CURRENT_TYPE, + MPI_COMM_WORLD); + } + #ifdef HAVE_TRACING TRACE_smpi_collective_out(rank, -1, __FUNCTION__); TRACE_smpi_computing_in(rank); #endif - XBT_VERB("%s %f", xbt_str_join_array(action, " "), - smpi_process_simulated_elapsed()-clock); + + log_timed_action (action, clock); + xbt_free(send); + xbt_free(recv); +} + +static void action_allToAllv(const char *const *action) { + /* + The structure of the allToAllV action for the rank 0 (total 4 processes) + is the following: + 0 allToAllV 100 1 7 10 12 5 10 20 45 100 1 70 10 5 1 5 77 90 + + where: + 1) 100 is the size of the send buffer *sizeof(int), + 2) 1 7 10 12 is the sendcounts array + 3) 5 10 20 45 is the sdispls array + 4) 100*sizeof(int) is the size of the receiver buffer + 5) 1 70 10 5 is the recvcounts array + 6) 1 5 77 90 is the rdispls array + + */ + + + double clock = smpi_process_simulated_elapsed(); + + int comm_size = smpi_comm_size(MPI_COMM_WORLD); + 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); + int *senddisps = xbt_new0(int, comm_size); + int *recvdisps = xbt_new0(int, comm_size); + + send_buf_size=parse_double(action[2]); + recv_buf_size=parse_double(action[3+2*comm_size]); + + int *sendbuf = xbt_new0(int, send_buf_size); + int *recvbuf = xbt_new0(int, recv_buf_size); + + if(action[4+4*comm_size]) MPI_CURRENT_TYPE=decode_datatype(action[4+4*comm_size]); + else MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; + + for(i=0;i