X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/b8ec45f42e58790c64cf449661f39f09f9b5fdb4..da41b147264c6eb7febc5be6cfb2380eab69ba70:/src/smpi/smpi_replay.c diff --git a/src/smpi/smpi_replay.c b/src/smpi/smpi_replay.c index 4811769e83..5991906ea4 100644 --- a/src/smpi/smpi_replay.c +++ b/src/smpi/smpi_replay.c @@ -10,13 +10,14 @@ #include #include -#define MPI_DTYPE MPI_BYTE - 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; +xbt_dynar_t *reqq = NULL; + +MPI_Datatype MPI_DEFAULT_TYPE; +MPI_Datatype MPI_CURRENT_TYPE; static void log_timed_action (const char *const *action, double clock){ if (XBT_LOG_ISENABLED(smpi_replay, xbt_log_priority_verbose)){ @@ -26,7 +27,6 @@ static void log_timed_action (const char *const *action, double clock){ } } - typedef struct { xbt_dynar_t isends; /* of MPI_Request */ xbt_dynar_t irecvs; /* of MPI_Request */ @@ -44,6 +44,40 @@ 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; @@ -51,7 +85,9 @@ static void action_init(const char *const *action) 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); @@ -60,20 +96,19 @@ static void action_init(const char *const *action) /*initialize the number of active processes */ active_processes = smpi_process_count(); - reqq=xbt_new0(xbt_dynar_t,active_processes); + if (!reqq) { + reqq=xbt_new0(xbt_dynar_t,active_processes); - for(i=0;iisends),xbt_dynar_length(globals->irecvs)); @@ -118,6 +153,13 @@ 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); @@ -126,7 +168,7 @@ static void action_send(const char *const *action) TRACE_smpi_send(rank, rank, dst_traced); #endif - smpi_mpi_send(NULL, size, MPI_DTYPE, to , 0, MPI_COMM_WORLD); + smpi_mpi_send(NULL, size, MPI_CURRENT_TYPE, to , 0, MPI_COMM_WORLD); log_timed_action (action, clock); @@ -143,6 +185,10 @@ static void action_Isend(const char *const *action) 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 @@ -153,7 +199,7 @@ static void action_Isend(const char *const *action) TRACE_smpi_send(rank, rank, dst_traced); #endif - request = smpi_mpi_isend(NULL, size, MPI_DTYPE, 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__); @@ -172,6 +218,10 @@ 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); @@ -180,7 +230,7 @@ static void action_recv(const char *const *action) { TRACE_smpi_ptp_in(rank, src_traced, rank, __FUNCTION__); #endif - smpi_mpi_recv(NULL, size, MPI_DTYPE, 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__); @@ -197,8 +247,12 @@ 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); @@ -206,7 +260,7 @@ 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_DTYPE, 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__); @@ -343,10 +397,25 @@ static void action_barrier(const char *const *action){ 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); @@ -354,7 +423,7 @@ static void action_bcast(const char *const *action) TRACE_smpi_collective_in(rank, root_traced, __FUNCTION__); #endif - smpi_mpi_bcast(NULL, size, MPI_DTYPE, 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); @@ -365,15 +434,27 @@ static void action_bcast(const char *const *action) 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_DTYPE, MPI_OP_NULL, 0, MPI_COMM_WORLD); + mpi_coll_reduce_fun(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); @@ -385,15 +466,19 @@ static void action_reduce(const char *const *action) 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_DTYPE, MPI_OP_NULL, 0, MPI_COMM_WORLD); + mpi_coll_reduce_fun(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_DTYPE, 0, MPI_COMM_WORLD); + 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); @@ -405,32 +490,84 @@ static void action_allReduce(const char *const *action) { 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 = atoi(action[2]); - int recv_size = atoi(action[3]); - void *send = xbt_new0(int, send_size*comm_size); - void *recv = xbt_new0(int, send_size*comm_size); - + int send_size = parse_double(action[2]); + int recv_size = parse_double(action[3]); + MPI_Datatype MPI_CURRENT_TYPE2; + + if(action[4]) { + MPI_CURRENT_TYPE=decode_datatype(action[4]); + MPI_CURRENT_TYPE2=decode_datatype(action[5]); + } + else { + 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)); #ifdef HAVE_TRACING int rank = smpi_process_index(); TRACE_smpi_computing_out(rank); TRACE_smpi_collective_in(rank, -1, __FUNCTION__); #endif + + mpi_coll_alltoall_fun(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, MPI_COMM_WORLD); - if (send_size < 200 && comm_size > 12) { - smpi_coll_tuned_alltoall_bruck(send, send_size, MPI_DTYPE, - recv, recv_size, MPI_DTYPE, - MPI_COMM_WORLD); - } else if (send_size < 3000) { +#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); +} + + +static void action_gather(const char *const *action) { - smpi_coll_tuned_alltoall_basic_linear(send, send_size, MPI_DTYPE, - recv, recv_size, MPI_DTYPE, - MPI_COMM_WORLD); + + /* + The structure of the gather action for the rank 0 (total 4 processes) + is the following: + 0 gather 68 68 0 0 0 + + where: + 1) 68 is the sendcounts + 2) 68 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); + int send_size = parse_double(action[2]); + int recv_size = parse_double(action[3]); + MPI_Datatype MPI_CURRENT_TYPE2; + if(action[5]) { + MPI_CURRENT_TYPE=decode_datatype(action[5]); + MPI_CURRENT_TYPE2=decode_datatype(action[6]); } else { - smpi_coll_tuned_alltoall_pairwise(send, send_size, MPI_DTYPE, - recv, recv_size, MPI_DTYPE, - MPI_COMM_WORLD); + 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(); + + if(rank==root) + recv = calloc(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__); +#endif +smpi_mpi_gather(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__); @@ -442,6 +579,126 @@ static void action_allToAll(const char *const *action) { xbt_free(recv); } + +static void action_reducescatter(const char *const *action) { + + /* + The structure of the reducescatter action for the rank 0 (total 4 processes) + is the following: +0 reduceScatter 275427 275427 275427 204020 11346849 0 + + where: + 1) The first four values after the name of the action declare the recvcounts array + 2) The value 11346849 is the amount of instructions + 3) The last value corresponds to the datatype, see decode_datatype(). + + We analyze a MPI_Reduce_scatter call to one MPI_Reduce and one MPI_Scatterv. + + */ + + double clock = smpi_process_simulated_elapsed(); + int comm_size = smpi_comm_size(MPI_COMM_WORLD); + 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 rank = smpi_process_index(); + + if(action[3+comm_size]) + MPI_CURRENT_TYPE=decode_datatype(action[3+comm_size]); + else + MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; + + for(i=0;i