int mca_fcoll_static_file_read_all (mca_io_ompio_file_t *fh, void *buf, int count, struct ompi_datatype_t *datatype, ompi_status_public_t *status) { int ret = OMPI_SUCCESS, iov_size=0, *bytes_remaining=NULL; int i, j, l,cycles=0, local_cycles=0, *current_index=NULL; int index, *disp_index=NULL, *bytes_per_process=NULL, current_position=0; int **blocklen_per_process=NULL, *iovec_count_per_process=NULL; int *displs=NULL, *sorted=NULL ,entries_per_aggregator=0; int *sorted_file_offsets=NULL, temp_index=0, position=0, *temp_disp_index=NULL; MPI_Aint **displs_per_process=NULL, global_iov_count=0, global_count=0; MPI_Aint *memory_displacements=NULL; int bytes_to_read_in_cycle=0; size_t max_data=0, bytes_per_cycle=0; uint32_t iov_count=0, iov_index=0; struct iovec *decoded_iov=NULL, *iov=NULL; mca_fcoll_static_local_io_array *local_iov_array=NULL, *global_iov_array=NULL; mca_fcoll_static_local_io_array *file_offsets_for_agg=NULL; char *global_buf=NULL, *receive_buf=NULL; int blocklen[3] = {1, 1, 1}; int static_num_io_procs=1; OPAL_PTRDIFF_TYPE d[3], base; ompi_datatype_t *types[3]; ompi_datatype_t *io_array_type=MPI_DATATYPE_NULL; ompi_datatype_t **sendtype = NULL; MPI_Request *send_req=NULL, recv_req=NULL; int my_aggregator=-1; bool recvbuf_is_contiguous=false; size_t ftype_size; OPAL_PTRDIFF_TYPE ftype_extent, lb; #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN double read_time = 0.0, start_read_time = 0.0, end_read_time = 0.0; double rcomm_time = 0.0, start_rcomm_time = 0.0, end_rcomm_time = 0.0; double read_exch = 0.0, start_rexch = 0.0, end_rexch = 0.0; mca_common_ompio_print_entry nentry; #endif #if DEBUG_ON MPI_Aint gc_in; #endif opal_datatype_type_size ( &datatype->super, &ftype_size ); opal_datatype_get_extent ( &datatype->super, &lb, &ftype_extent ); /************************************************************************** ** 1. In case the data is not contigous in memory, decode it into an iovec **************************************************************************/ if ( ( ftype_extent == (OPAL_PTRDIFF_TYPE) ftype_size) && opal_datatype_is_contiguous_memory_layout(&datatype->super,1) && 0 == lb ) { recvbuf_is_contiguous = true; } /* In case the data is not contigous in memory, decode it into an iovec */ if (!recvbuf_is_contiguous ) { fh->f_decode_datatype ( (struct mca_io_ompio_file_t *)fh, datatype, count, buf, &max_data, &decoded_iov, &iov_count); } else { max_data = count * datatype->super.size; } if ( MPI_STATUS_IGNORE != status ) { status->_ucount = max_data; } fh->f_get_num_aggregators ( &static_num_io_procs ); fh->f_set_aggregator_props ((struct mca_io_ompio_file_t *) fh, static_num_io_procs, max_data); my_aggregator = fh->f_procs_in_group[fh->f_aggregator_index]; /* printf("max_data %ld\n", max_data); */ ret = fh->f_generate_current_file_view((struct mca_io_ompio_file_t *)fh, max_data, &iov, &iov_size); if (ret != OMPI_SUCCESS){ goto exit; } if ( iov_size > 0 ) { local_iov_array = (mca_fcoll_static_local_io_array *)malloc (iov_size * sizeof(mca_fcoll_static_local_io_array)); if ( NULL == local_iov_array){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } for (j=0; j < iov_size; j++){ local_iov_array[j].offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t) iov[j].iov_base; local_iov_array[j].length = (size_t)iov[j].iov_len; local_iov_array[j].process_id = fh->f_rank; } } else { /* Allocate at least one element to correctly create the derived data type */ local_iov_array = (mca_fcoll_static_local_io_array *)malloc (sizeof(mca_fcoll_static_local_io_array)); if ( NULL == local_iov_array){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } local_iov_array[0].offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t) 0; local_iov_array[0].length = (size_t) 0; local_iov_array[0].process_id = fh->f_rank; } d[0] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0]; d[1] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0].length; d[2] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0].process_id; base = d[0]; for (i=0 ; i<3 ; i++) { d[i] -= base; } /* io_array datatype for using in communication*/ types[0] = &ompi_mpi_long.dt; types[1] = &ompi_mpi_long.dt; types[2] = &ompi_mpi_int.dt; ompi_datatype_create_struct (3, blocklen, d, types, &io_array_type); ompi_datatype_commit (&io_array_type); /* #########################################################*/ fh->f_get_bytes_per_agg ( (int*) &bytes_per_cycle); local_cycles = ceil((double)max_data*fh->f_procs_per_group/bytes_per_cycle); #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rexch = MPI_Wtime(); #endif ret = fh->f_comm->c_coll.coll_allreduce (&local_cycles, &cycles, 1, MPI_INT, MPI_MAX, fh->f_comm, fh->f_comm->c_coll.coll_allreduce_module); if (OMPI_SUCCESS != ret){ goto exit; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif if (my_aggregator == fh->f_rank) { disp_index = (int *) malloc (fh->f_procs_per_group * sizeof(int)); if (NULL == disp_index) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } bytes_per_process = (int *) malloc (fh->f_procs_per_group * sizeof(int )); if (NULL == bytes_per_process){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } bytes_remaining = (int *) calloc (fh->f_procs_per_group, sizeof(int)); if (NULL == bytes_remaining){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } current_index = (int *) calloc (fh->f_procs_per_group, sizeof(int)); if (NULL == current_index){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } blocklen_per_process = (int **)calloc (fh->f_procs_per_group, sizeof (int*)); if (NULL == blocklen_per_process) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } displs_per_process = (MPI_Aint **)calloc (fh->f_procs_per_group, sizeof (MPI_Aint*)); if (NULL == displs_per_process) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } } iovec_count_per_process = (int *) calloc (fh->f_procs_per_group, sizeof(int)); if (NULL == iovec_count_per_process){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } displs = (int *) calloc (fh->f_procs_per_group, sizeof(int)); if (NULL == displs){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rexch = MPI_Wtime(); #endif ret = fcoll_base_coll_allgather_array (&iov_size, 1, MPI_INT, iovec_count_per_process, 1, MPI_INT, fh->f_aggregator_index, fh->f_procs_in_group, fh->f_procs_per_group, fh->f_comm); if( OMPI_SUCCESS != ret){ goto exit; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif if (my_aggregator == fh->f_rank) { displs[0] = 0; global_iov_count = iovec_count_per_process[0]; for (i=1 ; i<fh->f_procs_per_group ; i++) { global_iov_count += iovec_count_per_process[i]; displs[i] = displs[i-1] + iovec_count_per_process[i-1]; } } if ( (my_aggregator == fh->f_rank) && (global_iov_count > 0 )) { global_iov_array = (mca_fcoll_static_local_io_array *) malloc (global_iov_count * sizeof(mca_fcoll_static_local_io_array)); if (NULL == global_iov_array){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rexch = MPI_Wtime(); #endif ret = fcoll_base_coll_gatherv_array (local_iov_array, iov_size, io_array_type, global_iov_array, iovec_count_per_process, displs, io_array_type, fh->f_aggregator_index, fh->f_procs_in_group, fh->f_procs_per_group, fh->f_comm); if (OMPI_SUCCESS != ret){ fprintf(stderr,"global_iov_array gather error!\n"); goto exit; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif if (NULL != local_iov_array){ free(local_iov_array); local_iov_array = NULL; } if ( ( my_aggregator == fh->f_rank) && ( global_iov_count > 0 )) { sorted = (int *)malloc (global_iov_count * sizeof(int)); if (NULL == sorted) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } read_local_heap_sort (global_iov_array, global_iov_count, sorted); send_req = (MPI_Request *) malloc (fh->f_procs_per_group * sizeof(MPI_Request)); if (NULL == send_req){ opal_output ( 1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } sendtype = (ompi_datatype_t **) malloc (fh->f_procs_per_group * sizeof(ompi_datatype_t *)); if (NULL == sendtype) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } for ( i=0; i<fh->f_procs_per_group; i++ ) { sendtype[i] = MPI_DATATYPE_NULL; } if (NULL == bytes_per_process){ bytes_per_process = (int *) malloc (fh->f_procs_per_group * sizeof(int)); if (NULL == bytes_per_process){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } } } #if DEBUG_ON if (my_aggregator == fh->f_rank) { for (gc_in=0; gc_in<global_iov_count; gc_in++){ printf("%d: Offset[%ld]: %lld, Length[%ld]: %ld\n", global_iov_array[sorted[gc_in]].process_id, gc_in, global_iov_array[sorted[gc_in]].offset, gc_in, global_iov_array[sorted[gc_in]].length); } } #endif #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rexch = MPI_Wtime(); #endif for (index = 0; index < cycles; index++){ if (my_aggregator == fh->f_rank) { fh->f_num_of_io_entries = 0; if (NULL != fh->f_io_array) { free (fh->f_io_array); fh->f_io_array = NULL; } if (NULL != global_buf) { free (global_buf); global_buf = NULL; } if (NULL != sorted_file_offsets){ free(sorted_file_offsets); sorted_file_offsets = NULL; } if (NULL != file_offsets_for_agg){ free(file_offsets_for_agg); file_offsets_for_agg = NULL; } if (NULL != memory_displacements){ free(memory_displacements); memory_displacements= NULL; } if ( NULL != sendtype ) { for ( i=0; i<fh->f_procs_per_group; i++ ) { if ( MPI_DATATYPE_NULL != sendtype[i] ) { ompi_datatype_destroy (&sendtype[i] ); sendtype[i] = MPI_DATATYPE_NULL; } } } for(l=0;l<fh->f_procs_per_group;l++){ disp_index[l] = 1; if (NULL != blocklen_per_process[l]){ free(blocklen_per_process[l]); blocklen_per_process[l] = NULL; } if (NULL != displs_per_process[l]){ free(displs_per_process[l]); displs_per_process[l] = NULL; } blocklen_per_process[l] = (int *) calloc (1, sizeof(int)); if (NULL == blocklen_per_process[l]) { opal_output (1, "OUT OF MEMORY for blocklen\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } displs_per_process[l] = (MPI_Aint *) calloc (1, sizeof(MPI_Aint)); if (NULL == displs_per_process[l]){ opal_output (1, "OUT OF MEMORY for displs\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } } } if (index < local_cycles ) { if ((index == local_cycles-1) && (max_data % (bytes_per_cycle/fh->f_procs_per_group))) { bytes_to_read_in_cycle = max_data - position; } else if (max_data <= bytes_per_cycle/fh->f_procs_per_group) { bytes_to_read_in_cycle = max_data; } else { bytes_to_read_in_cycle = bytes_per_cycle/fh->f_procs_per_group; } } else { bytes_to_read_in_cycle = 0; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rexch = MPI_Wtime(); #endif fcoll_base_coll_gather_array (&bytes_to_read_in_cycle, 1, MPI_INT, bytes_per_process, 1, MPI_INT, fh->f_aggregator_index, fh->f_procs_in_group, fh->f_procs_per_group, fh->f_comm); #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif if (recvbuf_is_contiguous ) { receive_buf = &((char*)buf)[position]; } else if (bytes_to_read_in_cycle) { receive_buf = (char *) malloc (bytes_to_read_in_cycle * sizeof(char)); if ( NULL == receive_buf){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rcomm_time = MPI_Wtime(); #endif ret = MCA_PML_CALL(irecv(receive_buf, bytes_to_read_in_cycle, MPI_BYTE, my_aggregator, 123, fh->f_comm, &recv_req)); if (OMPI_SUCCESS != ret){ goto exit; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif if (my_aggregator == fh->f_rank) { for (i=0;i<fh->f_procs_per_group; i++){ while (bytes_per_process[i] > 0){ /*printf("%d: bytes_per_process[%d]: %d, bytes_remaining[%d]: %d\n", index, i, bytes_per_process[i], i, bytes_remaining[i]);*/ if (read_get_process_id(global_iov_array[sorted[current_index[i]]].process_id, fh) == i){ /* current id owns this entry!*/ if (bytes_remaining[i]){ /*Remaining bytes in the current entry of the global offset array*/ if (bytes_remaining[i] <= bytes_per_process[i]){ blocklen_per_process[i][disp_index[i] - 1] = bytes_remaining[i]; displs_per_process[i][disp_index[i] - 1] = global_iov_array[sorted[current_index[i]]].offset + (global_iov_array[sorted[current_index[i]]].length - bytes_remaining[i]); blocklen_per_process[i] = (int *) realloc ((void *)blocklen_per_process[i], (disp_index[i]+1)*sizeof(int)); displs_per_process[i] = (MPI_Aint *)realloc ((void *)displs_per_process[i], (disp_index[i]+1)*sizeof(MPI_Aint)); bytes_per_process[i] -= bytes_remaining[i]; blocklen_per_process[i][disp_index[i]] = 0; displs_per_process[i][disp_index[i]] = 0; disp_index[i] += 1; bytes_remaining[i] = 0; /* This entry has been used up, we need to move to the next entry of this process and make current_index point there*/ current_index[i] = read_find_next_index(i, current_index[i], fh, global_iov_array, global_iov_count, sorted); if (current_index[i] == -1){ break; } continue; } else{ blocklen_per_process[i][disp_index[i] - 1] = bytes_per_process[i]; displs_per_process[i][disp_index[i] - 1] = global_iov_array[sorted[current_index[i]]].offset + (global_iov_array[sorted[current_index[i]]].length - bytes_remaining[i]); bytes_remaining[i] -= bytes_per_process[i]; bytes_per_process[i] = 0; break; } } else{ if (bytes_per_process[i] < global_iov_array[sorted[current_index[i]]].length){ blocklen_per_process[i][disp_index[i] - 1] = bytes_per_process[i]; displs_per_process[i][disp_index[i] - 1] = global_iov_array[sorted[current_index[i]]].offset; bytes_remaining[i] = global_iov_array[sorted[current_index[i]]].length - bytes_per_process[i]; bytes_per_process[i] = 0; break; } else { blocklen_per_process[i][disp_index[i] - 1] = global_iov_array[sorted[current_index[i]]].length; displs_per_process[i][disp_index[i] - 1] = global_iov_array[sorted[current_index[i]]].offset; blocklen_per_process[i] = (int *) realloc ((void *)blocklen_per_process[i], (disp_index[i]+1)*sizeof(int)); displs_per_process[i] = (MPI_Aint *)realloc ((void *)displs_per_process[i], (disp_index[i]+1)*sizeof(MPI_Aint)); blocklen_per_process[i][disp_index[i]] = 0; displs_per_process[i][disp_index[i]] = 0; disp_index[i] += 1; bytes_per_process[i] -= global_iov_array[sorted[current_index[i]]].length; current_index[i] = read_find_next_index(i, current_index[i], fh, global_iov_array, global_iov_count, sorted); if (current_index[i] == -1){ break; } } } } else{ current_index[i] = read_find_next_index(i, current_index[i], fh, global_iov_array, global_iov_count, sorted); if (current_index[i] == -1){ bytes_per_process[i] = 0; /* no more entries left to service this request*/ continue; } } } } entries_per_aggregator=0; for (i=0;i<fh->f_procs_per_group;i++){ for (j=0;j<disp_index[i];j++){ if (blocklen_per_process[i][j] > 0){ entries_per_aggregator++; #if DEBUG_ON printf("%d sends blocklen[%d]: %d, disp[%d]: %ld to %d\n", fh->f_procs_in_group[i],j, blocklen_per_process[i][j],j, displs_per_process[i][j], fh->f_rank); #endif } } } if (entries_per_aggregator > 0){ file_offsets_for_agg = (mca_fcoll_static_local_io_array *) malloc(entries_per_aggregator*sizeof(mca_fcoll_static_local_io_array)); if (NULL == file_offsets_for_agg) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } sorted_file_offsets = (int *) malloc (entries_per_aggregator * sizeof(int)); if (NULL == sorted_file_offsets){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } temp_index=0; global_count = 0; for (i=0;i<fh->f_procs_per_group; i++){ for(j=0;j<disp_index[i]; j++){ if (blocklen_per_process[i][j] > 0){ file_offsets_for_agg[temp_index].length = blocklen_per_process[i][j]; global_count += blocklen_per_process[i][j]; file_offsets_for_agg[temp_index].process_id = i; file_offsets_for_agg[temp_index].offset = displs_per_process[i][j]; temp_index++; } } } } else{ continue; } read_local_heap_sort (file_offsets_for_agg, entries_per_aggregator, sorted_file_offsets); memory_displacements = (MPI_Aint *) malloc (entries_per_aggregator * sizeof(MPI_Aint)); memory_displacements[sorted_file_offsets[0]] = 0; for (i=1; i<entries_per_aggregator; i++){ memory_displacements[sorted_file_offsets[i]] = memory_displacements[sorted_file_offsets[i-1]] + file_offsets_for_agg[sorted_file_offsets[i-1]].length; } global_buf = (char *) malloc (global_count * sizeof(char)); if (NULL == global_buf){ opal_output(1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } #if DEBUG_ON printf("************Cycle: %d, Aggregator: %d ***************\n", index+1,fh->f_rank); for (i=0; i<entries_per_aggregator;i++){ printf("%d: OFFSET: %lld LENGTH: %ld, Mem-offset: %ld, disp_index :%d\n", file_offsets_for_agg[sorted_file_offsets[i]].process_id, file_offsets_for_agg[sorted_file_offsets[i]].offset, file_offsets_for_agg[sorted_file_offsets[i]].length, memory_displacements[sorted_file_offsets[i]], disp_index[i]); } #endif fh->f_io_array = (mca_io_ompio_io_array_t *) malloc (entries_per_aggregator * sizeof (mca_io_ompio_io_array_t)); if (NULL == fh->f_io_array) { opal_output(1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } fh->f_num_of_io_entries = 0; fh->f_io_array[0].offset = (IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[0]].offset; fh->f_io_array[0].length = file_offsets_for_agg[sorted_file_offsets[0]].length; fh->f_io_array[0].memory_address = global_buf+memory_displacements[sorted_file_offsets[0]]; fh->f_num_of_io_entries++; for (i=1;i<entries_per_aggregator;i++){ if (file_offsets_for_agg[sorted_file_offsets[i-1]].offset + file_offsets_for_agg[sorted_file_offsets[i-1]].length == file_offsets_for_agg[sorted_file_offsets[i]].offset){ fh->f_io_array[fh->f_num_of_io_entries - 1].length += file_offsets_for_agg[sorted_file_offsets[i]].length; } else{ fh->f_io_array[fh->f_num_of_io_entries].offset = (IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[i]].offset; fh->f_io_array[fh->f_num_of_io_entries].length = file_offsets_for_agg[sorted_file_offsets[i]].length; fh->f_io_array[fh->f_num_of_io_entries].memory_address = global_buf+memory_displacements[sorted_file_offsets[i]]; fh->f_num_of_io_entries++; } } #if DEBUG_ON printf("*************************** %d\n", fh->f_num_of_io_entries); for (i=0 ; i<fh->f_num_of_io_entries ; i++) { printf(" ADDRESS: %p OFFSET: %ld LENGTH: %ld\n", fh->f_io_array[i].memory_address, (OPAL_PTRDIFF_TYPE)fh->f_io_array[i].offset, fh->f_io_array[i].length); } #endif #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_read_time = MPI_Wtime(); #endif if (fh->f_num_of_io_entries) { if ( 0 > fh->f_fbtl->fbtl_preadv (fh)) { opal_output (1, "READ FAILED\n"); ret = OMPI_ERROR; goto exit; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_read_time = MPI_Wtime(); read_time += end_read_time - start_read_time; #endif #if DEBUG_ON printf("************Cycle: %d, Aggregator: %d ***************\n", index+1,fh->f_rank); if (my_aggregator == fh->f_rank){ for (i=0 ; i<global_count/4 ; i++) printf (" READ %d \n",((int *)global_buf)[i]); } #endif temp_disp_index = (int *)calloc (1, fh->f_procs_per_group * sizeof (int)); if (NULL == temp_disp_index) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } for (i=0; i<entries_per_aggregator; i++){ temp_index = file_offsets_for_agg[sorted_file_offsets[i]].process_id; displs_per_process[temp_index][temp_disp_index[temp_index]] = memory_displacements[sorted_file_offsets[i]]; if (temp_disp_index[temp_index] < disp_index[temp_index]){ temp_disp_index[temp_index] += 1; } else{ printf("temp_disp_index[%d]: %d is greater than disp_index[%d]: %d\n", temp_index, temp_disp_index[temp_index], temp_index, disp_index[temp_index]); } } if (NULL != temp_disp_index){ free(temp_disp_index); temp_disp_index = NULL; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rcomm_time = MPI_Wtime(); #endif for (i=0;i<fh->f_procs_per_group; i++){ send_req[i] = MPI_REQUEST_NULL; ompi_datatype_create_hindexed(disp_index[i], blocklen_per_process[i], displs_per_process[i], MPI_BYTE, &sendtype[i]); ompi_datatype_commit(&sendtype[i]); ret = MCA_PML_CALL (isend(global_buf, 1, sendtype[i], fh->f_procs_in_group[i], 123, MCA_PML_BASE_SEND_STANDARD, fh->f_comm, &send_req[i])); if(OMPI_SUCCESS != ret){ goto exit; } } ret = ompi_request_wait_all (fh->f_procs_per_group, send_req, MPI_STATUS_IGNORE); if (OMPI_SUCCESS != ret){ goto exit; } } /* if ( my_aggregator == fh->f_rank ) */ ret = ompi_request_wait (&recv_req, MPI_STATUS_IGNORE); if (OMPI_SUCCESS != ret){ goto exit; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif position += bytes_to_read_in_cycle; if (!recvbuf_is_contiguous) { OPAL_PTRDIFF_TYPE mem_address; size_t remaining = 0; size_t temp_position = 0; remaining = bytes_to_read_in_cycle; while (remaining && (iov_count > iov_index)){ mem_address = (OPAL_PTRDIFF_TYPE) (decoded_iov[iov_index].iov_base) + current_position; if (remaining >= (decoded_iov[iov_index].iov_len - current_position)) { memcpy ((IOVBASE_TYPE *) mem_address, receive_buf+temp_position, decoded_iov[iov_index].iov_len - current_position); remaining = remaining - (decoded_iov[iov_index].iov_len - current_position); temp_position = temp_position + (decoded_iov[iov_index].iov_len - current_position); iov_index = iov_index + 1; current_position = 0; } else{ memcpy ((IOVBASE_TYPE *) mem_address, receive_buf+temp_position, remaining); current_position = current_position + remaining; remaining = 0; } } if (NULL != receive_buf) { free (receive_buf); receive_buf = NULL; } } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rexch = MPI_Wtime(); read_exch += end_rexch - start_rexch; nentry.time[0] = read_time; nentry.time[1] = rcomm_time; nentry.time[2] = read_exch; if (my_aggregator == fh->f_rank) nentry.aggregator = 1; else nentry.aggregator = 0; nentry.nprocs_for_coll = static_num_io_procs; if (!mca_common_ompio_full_print_queue(fh->f_coll_read_time)){ mca_common_ompio_register_print_entry(fh->f_coll_read_time, nentry); } #endif exit: if (NULL != decoded_iov){ free(decoded_iov); decoded_iov = NULL; } if (NULL != displs){ free(displs); displs = NULL; } if (NULL != iovec_count_per_process){ free(iovec_count_per_process); iovec_count_per_process=NULL; } if (NULL != local_iov_array){ free(local_iov_array); local_iov_array=NULL; } if (NULL != global_iov_array){ free(global_iov_array); global_iov_array=NULL; } if (my_aggregator == fh->f_rank) { for(l=0;l<fh->f_procs_per_group;l++){ if (blocklen_per_process) { free(blocklen_per_process[l]); } if (NULL != displs_per_process[l]){ free(displs_per_process[l]); displs_per_process[l] = NULL; } } } if (NULL != bytes_per_process){ free(bytes_per_process); bytes_per_process =NULL; } if (NULL != disp_index){ free(disp_index); disp_index =NULL; } if (NULL != displs_per_process){ free(displs_per_process); displs_per_process = NULL; } if(NULL != bytes_remaining){ free(bytes_remaining); bytes_remaining = NULL; } if(NULL != current_index){ free(current_index); current_index = NULL; } if (NULL != blocklen_per_process){ free(blocklen_per_process); blocklen_per_process =NULL; } if (NULL != bytes_remaining){ free(bytes_remaining); bytes_remaining =NULL; } if (NULL != memory_displacements){ free(memory_displacements); memory_displacements= NULL; } if (NULL != file_offsets_for_agg){ free(file_offsets_for_agg); file_offsets_for_agg = NULL; } if (NULL != sorted_file_offsets){ free(sorted_file_offsets); sorted_file_offsets = NULL; } if (NULL != sendtype){ free(sendtype); sendtype=NULL; } if ( !recvbuf_is_contiguous ) { if (NULL != receive_buf){ free(receive_buf); receive_buf=NULL; } } if (NULL != global_buf) { free(global_buf); global_buf = NULL; } if (NULL != sorted) { free(sorted); sorted = NULL; } if (NULL != send_req){ free(send_req); send_req = NULL; } return ret; }
int mca_fcoll_vulcan_file_read_all (ompio_file_t *fh, void *buf, int count, struct ompi_datatype_t *datatype, ompi_status_public_t *status) { MPI_Aint position = 0; MPI_Aint total_bytes = 0; /* total bytes to be read */ MPI_Aint bytes_to_read_in_cycle = 0; /* left to be read in a cycle*/ MPI_Aint bytes_per_cycle = 0; /* total read in each cycle by each process*/ int index = 0, ret=OMPI_SUCCESS; int cycles = 0; int i=0, j=0, l=0; int n=0; /* current position in total_bytes_per_process array */ MPI_Aint bytes_remaining = 0; /* how many bytes have been read from the current value from total_bytes_per_process */ int *sorted_file_offsets=NULL, entries_per_aggregator=0; int bytes_received = 0; int blocks = 0; /* iovec structure and count of the buffer passed in */ uint32_t iov_count = 0; struct iovec *decoded_iov = NULL; int iov_index = 0; size_t current_position = 0; struct iovec *local_iov_array=NULL, *global_iov_array=NULL; char *receive_buf = NULL; MPI_Aint *memory_displacements=NULL; /* global iovec at the readers that contain the iovecs created from file_set_view */ uint32_t total_fview_count = 0; int local_count = 0; int *fview_count = NULL, *disp_index=NULL, *temp_disp_index=NULL; int current_index=0, temp_index=0; int **blocklen_per_process=NULL; MPI_Aint **displs_per_process=NULL; char *global_buf = NULL; MPI_Aint global_count = 0; mca_io_ompio_local_io_array *file_offsets_for_agg=NULL; /* array that contains the sorted indices of the global_iov */ int *sorted = NULL; int *displs = NULL; int vulcan_num_io_procs; size_t max_data = 0; MPI_Aint *total_bytes_per_process = NULL; ompi_datatype_t **sendtype = NULL; MPI_Request *send_req=NULL, recv_req=NULL; int my_aggregator =-1; bool recvbuf_is_contiguous=false; size_t ftype_size; ptrdiff_t ftype_extent, lb; #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN double read_time = 0.0, start_read_time = 0.0, end_read_time = 0.0; double rcomm_time = 0.0, start_rcomm_time = 0.0, end_rcomm_time = 0.0; double read_exch = 0.0, start_rexch = 0.0, end_rexch = 0.0; mca_common_ompio_print_entry nentry; #endif /************************************************************************** ** 1. In case the data is not contigous in memory, decode it into an iovec **************************************************************************/ opal_datatype_type_size ( &datatype->super, &ftype_size ); opal_datatype_get_extent ( &datatype->super, &lb, &ftype_extent ); if ( (ftype_extent == (ptrdiff_t) ftype_size) && opal_datatype_is_contiguous_memory_layout(&datatype->super,1) && 0 == lb ) { recvbuf_is_contiguous = true; } if (! recvbuf_is_contiguous ) { ret = mca_common_ompio_decode_datatype ((struct ompio_file_t *)fh, datatype, count, buf, &max_data, &decoded_iov, &iov_count); if (OMPI_SUCCESS != ret){ goto exit; } } else { max_data = count * datatype->super.size; } if ( MPI_STATUS_IGNORE != status ) { status->_ucount = max_data; } vulcan_num_io_procs = fh->f_get_mca_parameter_value ( "num_aggregators", strlen ("num_aggregators")); if ( OMPI_ERR_MAX == vulcan_num_io_procs ) { ret = OMPI_ERROR; goto exit; } ret = mca_common_ompio_set_aggregator_props ((struct ompio_file_t *) fh, vulcan_num_io_procs, max_data); if (OMPI_SUCCESS != ret){ goto exit; } my_aggregator = fh->f_procs_in_group[0]; /************************************************************************** ** 2. Determine the total amount of data to be written **************************************************************************/ total_bytes_per_process = (MPI_Aint*)malloc(fh->f_procs_per_group*sizeof(MPI_Aint)); if (NULL == total_bytes_per_process) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rcomm_time = MPI_Wtime(); #endif ret = ompi_fcoll_base_coll_allgather_array (&max_data, 1, MPI_LONG, total_bytes_per_process, 1, MPI_LONG, 0, fh->f_procs_in_group, fh->f_procs_per_group, fh->f_comm); if (OMPI_SUCCESS != ret){ goto exit; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif for (i=0 ; i<fh->f_procs_per_group ; i++) { total_bytes += total_bytes_per_process[i]; } if (NULL != total_bytes_per_process) { free (total_bytes_per_process); total_bytes_per_process = NULL; } /********************************************************************* *** 3. Generate the File offsets/lengths corresponding to this write ********************************************************************/ ret = fh->f_generate_current_file_view ((struct ompio_file_t *) fh, max_data, &local_iov_array, &local_count); if (ret != OMPI_SUCCESS){ goto exit; } /************************************************************* *** 4. Allgather the File View information at all processes *************************************************************/ fview_count = (int *) malloc (fh->f_procs_per_group * sizeof (int)); if (NULL == fview_count) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rcomm_time = MPI_Wtime(); #endif ret = ompi_fcoll_base_coll_allgather_array (&local_count, 1, MPI_INT, fview_count, 1, MPI_INT, 0, fh->f_procs_in_group, fh->f_procs_per_group, fh->f_comm); if (OMPI_SUCCESS != ret){ goto exit; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif displs = (int*)malloc (fh->f_procs_per_group*sizeof(int)); if (NULL == displs) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } displs[0] = 0; total_fview_count = fview_count[0]; for (i=1 ; i<fh->f_procs_per_group ; i++) { total_fview_count += fview_count[i]; displs[i] = displs[i-1] + fview_count[i-1]; } #if DEBUG_ON if (my_aggregator == fh->f_rank) { for (i=0 ; i<fh->f_procs_per_group ; i++) { printf ("%d: PROCESS: %d ELEMENTS: %d DISPLS: %d\n", fh->f_rank, i, fview_count[i], displs[i]); } } #endif /* allocate the global iovec */ if (0 != total_fview_count) { global_iov_array = (struct iovec*)malloc (total_fview_count * sizeof(struct iovec)); if (NULL == global_iov_array) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rcomm_time = MPI_Wtime(); #endif ret = ompi_fcoll_base_coll_allgatherv_array (local_iov_array, local_count, fh->f_iov_type, global_iov_array, fview_count, displs, fh->f_iov_type, 0, fh->f_procs_in_group, fh->f_procs_per_group, fh->f_comm); if (OMPI_SUCCESS != ret){ goto exit; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif /**************************************************************************************** *** 5. Sort the global offset/lengths list based on the offsets. *** The result of the sort operation is the 'sorted', an integer array, *** which contains the indexes of the global_iov_array based on the offset. *** For example, if global_iov_array[x].offset is followed by global_iov_array[y].offset *** in the file, and that one is followed by global_iov_array[z].offset, than *** sorted[0] = x, sorted[1]=y and sorted[2]=z; ******************************************************************************************/ if (0 != total_fview_count) { sorted = (int *)malloc (total_fview_count * sizeof(int)); if (NULL == sorted) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } ompi_fcoll_base_sort_iovec (global_iov_array, total_fview_count, sorted); } if (NULL != local_iov_array) { free (local_iov_array); local_iov_array = NULL; } #if DEBUG_ON if (my_aggregator == fh->f_rank) { for (i=0 ; i<total_fview_count ; i++) { printf("%d: OFFSET: %p LENGTH: %d\n", fh->f_rank, global_iov_array[sorted[i]].iov_base, global_iov_array[sorted[i]].iov_len); } } #endif /************************************************************* *** 6. Determine the number of cycles required to execute this *** operation *************************************************************/ bytes_per_cycle = fh->f_get_mca_parameter_value ("bytes_per_agg", strlen ("bytes_per_agg")); if ( OMPI_ERR_MAX == bytes_per_cycle ) { ret = OMPI_ERROR; goto exit; } cycles = ceil((double)total_bytes/bytes_per_cycle); if ( my_aggregator == fh->f_rank) { disp_index = (int *)malloc (fh->f_procs_per_group * sizeof (int)); if (NULL == disp_index) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } blocklen_per_process = (int **)malloc (fh->f_procs_per_group * sizeof (int*)); if (NULL == blocklen_per_process) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } displs_per_process = (MPI_Aint **)malloc (fh->f_procs_per_group * sizeof (MPI_Aint*)); if (NULL == displs_per_process){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } for (i=0;i<fh->f_procs_per_group;i++){ blocklen_per_process[i] = NULL; displs_per_process[i] = NULL; } send_req = (MPI_Request *) malloc (fh->f_procs_per_group * sizeof(MPI_Request)); if (NULL == send_req){ opal_output ( 1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } global_buf = (char *) malloc (bytes_per_cycle); if (NULL == global_buf){ opal_output(1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } sendtype = (ompi_datatype_t **) malloc (fh->f_procs_per_group * sizeof(ompi_datatype_t *)); if (NULL == sendtype) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } for(l=0;l<fh->f_procs_per_group;l++){ sendtype[l] = MPI_DATATYPE_NULL; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rexch = MPI_Wtime(); #endif n = 0; bytes_remaining = 0; current_index = 0; for (index = 0; index < cycles; index++) { /********************************************************************** *** 7a. Getting ready for next cycle: initializing and freeing buffers **********************************************************************/ if (my_aggregator == fh->f_rank) { if (NULL != fh->f_io_array) { free (fh->f_io_array); fh->f_io_array = NULL; } fh->f_num_of_io_entries = 0; if (NULL != sendtype){ for (i =0; i< fh->f_procs_per_group; i++) { if ( MPI_DATATYPE_NULL != sendtype[i] ) { ompi_datatype_destroy(&sendtype[i]); sendtype[i] = MPI_DATATYPE_NULL; } } } for(l=0;l<fh->f_procs_per_group;l++){ disp_index[l] = 1; if (NULL != blocklen_per_process[l]){ free(blocklen_per_process[l]); blocklen_per_process[l] = NULL; } if (NULL != displs_per_process[l]){ free(displs_per_process[l]); displs_per_process[l] = NULL; } blocklen_per_process[l] = (int *) calloc (1, sizeof(int)); if (NULL == blocklen_per_process[l]) { opal_output (1, "OUT OF MEMORY for blocklen\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } displs_per_process[l] = (MPI_Aint *) calloc (1, sizeof(MPI_Aint)); if (NULL == displs_per_process[l]){ opal_output (1, "OUT OF MEMORY for displs\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } } if (NULL != sorted_file_offsets){ free(sorted_file_offsets); sorted_file_offsets = NULL; } if(NULL != file_offsets_for_agg){ free(file_offsets_for_agg); file_offsets_for_agg = NULL; } if (NULL != memory_displacements){ free(memory_displacements); memory_displacements = NULL; } } /* (my_aggregator == fh->f_rank */ /************************************************************************** *** 7b. Determine the number of bytes to be actually read in this cycle **************************************************************************/ if (cycles-1 == index) { bytes_to_read_in_cycle = total_bytes - bytes_per_cycle*index; } else { bytes_to_read_in_cycle = bytes_per_cycle; } #if DEBUG_ON if (my_aggregator == fh->f_rank) { printf ("****%d: CYCLE %d Bytes %d**********\n", fh->f_rank, index, bytes_to_write_in_cycle); } #endif /***************************************************************** *** 7c. Calculate how much data will be contributed in this cycle *** by each process *****************************************************************/ bytes_received = 0; while (bytes_to_read_in_cycle) { /* This next block identifies which process is the holder ** of the sorted[current_index] element; */ blocks = fview_count[0]; for (j=0 ; j<fh->f_procs_per_group ; j++) { if (sorted[current_index] < blocks) { n = j; break; } else { blocks += fview_count[j+1]; } } if (bytes_remaining) { /* Finish up a partially used buffer from the previous cycle */ if (bytes_remaining <= bytes_to_read_in_cycle) { /* Data fits completely into the block */ if (my_aggregator == fh->f_rank) { blocklen_per_process[n][disp_index[n] - 1] = bytes_remaining; displs_per_process[n][disp_index[n] - 1] = (ptrdiff_t)global_iov_array[sorted[current_index]].iov_base + (global_iov_array[sorted[current_index]].iov_len - bytes_remaining); blocklen_per_process[n] = (int *) realloc ((void *)blocklen_per_process[n], (disp_index[n]+1)*sizeof(int)); displs_per_process[n] = (MPI_Aint *) realloc ((void *)displs_per_process[n], (disp_index[n]+1)*sizeof(MPI_Aint)); blocklen_per_process[n][disp_index[n]] = 0; displs_per_process[n][disp_index[n]] = 0; disp_index[n] += 1; } if (fh->f_procs_in_group[n] == fh->f_rank) { bytes_received += bytes_remaining; } current_index ++; bytes_to_read_in_cycle -= bytes_remaining; bytes_remaining = 0; continue; } else { /* the remaining data from the previous cycle is larger than the bytes_to_write_in_cycle, so we have to segment again */ if (my_aggregator == fh->f_rank) { blocklen_per_process[n][disp_index[n] - 1] = bytes_to_read_in_cycle; displs_per_process[n][disp_index[n] - 1] = (ptrdiff_t)global_iov_array[sorted[current_index]].iov_base + (global_iov_array[sorted[current_index]].iov_len - bytes_remaining); } if (fh->f_procs_in_group[n] == fh->f_rank) { bytes_received += bytes_to_read_in_cycle; } bytes_remaining -= bytes_to_read_in_cycle; bytes_to_read_in_cycle = 0; break; } } else { /* No partially used entry available, have to start a new one */ if (bytes_to_read_in_cycle < (MPI_Aint) global_iov_array[sorted[current_index]].iov_len) { /* This entry has more data than we can sendin one cycle */ if (my_aggregator == fh->f_rank) { blocklen_per_process[n][disp_index[n] - 1] = bytes_to_read_in_cycle; displs_per_process[n][disp_index[n] - 1] = (ptrdiff_t)global_iov_array[sorted[current_index]].iov_base ; } if (fh->f_procs_in_group[n] == fh->f_rank) { bytes_received += bytes_to_read_in_cycle; } bytes_remaining = global_iov_array[sorted[current_index]].iov_len - bytes_to_read_in_cycle; bytes_to_read_in_cycle = 0; break; } else { /* Next data entry is less than bytes_to_write_in_cycle */ if (my_aggregator == fh->f_rank) { blocklen_per_process[n][disp_index[n] - 1] = global_iov_array[sorted[current_index]].iov_len; displs_per_process[n][disp_index[n] - 1] = (ptrdiff_t) global_iov_array[sorted[current_index]].iov_base; blocklen_per_process[n] = (int *) realloc ((void *)blocklen_per_process[n], (disp_index[n]+1)*sizeof(int)); displs_per_process[n] = (MPI_Aint *)realloc ((void *)displs_per_process[n], (disp_index[n]+1)*sizeof(MPI_Aint)); blocklen_per_process[n][disp_index[n]] = 0; displs_per_process[n][disp_index[n]] = 0; disp_index[n] += 1; } if (fh->f_procs_in_group[n] == fh->f_rank) { bytes_received += global_iov_array[sorted[current_index]].iov_len; } bytes_to_read_in_cycle -= global_iov_array[sorted[current_index]].iov_len; current_index ++; continue; } } } /* end while (bytes_to_read_in_cycle) */ /************************************************************************* *** 7d. Calculate the displacement on where to put the data and allocate *** the recieve buffer (global_buf) *************************************************************************/ if (my_aggregator == fh->f_rank) { entries_per_aggregator=0; for (i=0;i<fh->f_procs_per_group; i++){ for (j=0;j<disp_index[i];j++){ if (blocklen_per_process[i][j] > 0) entries_per_aggregator++ ; } } if (entries_per_aggregator > 0){ file_offsets_for_agg = (mca_io_ompio_local_io_array *) malloc(entries_per_aggregator*sizeof(mca_io_ompio_local_io_array)); if (NULL == file_offsets_for_agg) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } sorted_file_offsets = (int *) malloc (entries_per_aggregator*sizeof(int)); if (NULL == sorted_file_offsets){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } /*Moving file offsets to an IO array!*/ temp_index = 0; global_count = 0; for (i=0;i<fh->f_procs_per_group; i++){ for(j=0;j<disp_index[i];j++){ if (blocklen_per_process[i][j] > 0){ file_offsets_for_agg[temp_index].length = blocklen_per_process[i][j]; global_count += blocklen_per_process[i][j]; file_offsets_for_agg[temp_index].process_id = i; file_offsets_for_agg[temp_index].offset = displs_per_process[i][j]; temp_index++; } } } } else{ continue; } /* Sort the displacements for each aggregator */ read_heap_sort (file_offsets_for_agg, entries_per_aggregator, sorted_file_offsets); memory_displacements = (MPI_Aint *) malloc (entries_per_aggregator * sizeof(MPI_Aint)); memory_displacements[sorted_file_offsets[0]] = 0; for (i=1; i<entries_per_aggregator; i++){ memory_displacements[sorted_file_offsets[i]] = memory_displacements[sorted_file_offsets[i-1]] + file_offsets_for_agg[sorted_file_offsets[i-1]].length; } /********************************************************** *** 7e. Create the io array, and pass it to fbtl *********************************************************/ fh->f_io_array = (mca_common_ompio_io_array_t *) malloc (entries_per_aggregator * sizeof (mca_common_ompio_io_array_t)); if (NULL == fh->f_io_array) { opal_output(1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } fh->f_num_of_io_entries = 0; fh->f_io_array[0].offset = (IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[0]].offset; fh->f_io_array[0].length = file_offsets_for_agg[sorted_file_offsets[0]].length; fh->f_io_array[0].memory_address = global_buf+memory_displacements[sorted_file_offsets[0]]; fh->f_num_of_io_entries++; for (i=1;i<entries_per_aggregator;i++){ if (file_offsets_for_agg[sorted_file_offsets[i-1]].offset + file_offsets_for_agg[sorted_file_offsets[i-1]].length == file_offsets_for_agg[sorted_file_offsets[i]].offset){ fh->f_io_array[fh->f_num_of_io_entries - 1].length += file_offsets_for_agg[sorted_file_offsets[i]].length; } else{ fh->f_io_array[fh->f_num_of_io_entries].offset = (IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[i]].offset; fh->f_io_array[fh->f_num_of_io_entries].length = file_offsets_for_agg[sorted_file_offsets[i]].length; fh->f_io_array[fh->f_num_of_io_entries].memory_address = global_buf+memory_displacements[sorted_file_offsets[i]]; fh->f_num_of_io_entries++; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_read_time = MPI_Wtime(); #endif if (fh->f_num_of_io_entries) { if ( 0 > fh->f_fbtl->fbtl_preadv (fh)) { opal_output (1, "READ FAILED\n"); ret = OMPI_ERROR; goto exit; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_read_time = MPI_Wtime(); read_time += end_read_time - start_read_time; #endif /********************************************************** ******************** DONE READING ************************ *********************************************************/ temp_disp_index = (int *)calloc (1, fh->f_procs_per_group * sizeof (int)); if (NULL == temp_disp_index) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } for (i=0; i<entries_per_aggregator; i++){ temp_index = file_offsets_for_agg[sorted_file_offsets[i]].process_id; displs_per_process[temp_index][temp_disp_index[temp_index]] = memory_displacements[sorted_file_offsets[i]]; if (temp_disp_index[temp_index] < disp_index[temp_index]){ temp_disp_index[temp_index] += 1; } else{ printf("temp_disp_index[%d]: %d is greater than disp_index[%d]: %d\n", temp_index, temp_disp_index[temp_index], temp_index, disp_index[temp_index]); } } if (NULL != temp_disp_index){ free(temp_disp_index); temp_disp_index = NULL; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rcomm_time = MPI_Wtime(); #endif for (i=0;i<fh->f_procs_per_group;i++){ send_req[i] = MPI_REQUEST_NULL; if ( 0 < disp_index[i] ) { ompi_datatype_create_hindexed(disp_index[i], blocklen_per_process[i], displs_per_process[i], MPI_BYTE, &sendtype[i]); ompi_datatype_commit(&sendtype[i]); ret = MCA_PML_CALL (isend(global_buf, 1, sendtype[i], fh->f_procs_in_group[i], 123, MCA_PML_BASE_SEND_STANDARD, fh->f_comm, &send_req[i])); if(OMPI_SUCCESS != ret){ goto exit; } } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif } /********************************************************** *** 7f. Scatter the Data from the readers *********************************************************/ if ( recvbuf_is_contiguous ) { receive_buf = &((char*)buf)[position]; } else if (bytes_received) { /* allocate a receive buffer and copy the data that needs to be received into it in case the data is non-contigous in memory */ receive_buf = malloc (bytes_received); if (NULL == receive_buf) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rcomm_time = MPI_Wtime(); #endif ret = MCA_PML_CALL(irecv(receive_buf, bytes_received, MPI_BYTE, my_aggregator, 123, fh->f_comm, &recv_req)); if (OMPI_SUCCESS != ret){ goto exit; } if (my_aggregator == fh->f_rank){ ret = ompi_request_wait_all (fh->f_procs_per_group, send_req, MPI_STATUS_IGNORE); if (OMPI_SUCCESS != ret){ goto exit; } } ret = ompi_request_wait (&recv_req, MPI_STATUS_IGNORE); if (OMPI_SUCCESS != ret){ goto exit; } position += bytes_received; /* If data is not contigous in memory, copy the data from the receive buffer into the buffer passed in */ if (!recvbuf_is_contiguous ) { ptrdiff_t mem_address; size_t remaining = 0; size_t temp_position = 0; remaining = bytes_received; while (remaining) { mem_address = (ptrdiff_t) (decoded_iov[iov_index].iov_base) + current_position; if (remaining >= (decoded_iov[iov_index].iov_len - current_position)) { memcpy ((IOVBASE_TYPE *) mem_address, receive_buf+temp_position, decoded_iov[iov_index].iov_len - current_position); remaining = remaining - (decoded_iov[iov_index].iov_len - current_position); temp_position = temp_position + (decoded_iov[iov_index].iov_len - current_position); iov_index = iov_index + 1; current_position = 0; } else { memcpy ((IOVBASE_TYPE *) mem_address, receive_buf+temp_position, remaining); current_position = current_position + remaining; remaining = 0; } } if (NULL != receive_buf) { free (receive_buf); receive_buf = NULL; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif } /* end for (index=0; index < cycles; index ++) */ #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rexch = MPI_Wtime(); read_exch += end_rexch - start_rexch; nentry.time[0] = read_time; nentry.time[1] = rcomm_time; nentry.time[2] = read_exch; if (my_aggregator == fh->f_rank) nentry.aggregator = 1; else nentry.aggregator = 0; nentry.nprocs_for_coll = vulcan_num_io_procs; if (!mca_common_ompio_full_print_queue(fh->f_coll_read_time)){ mca_common_ompio_register_print_entry(fh->f_coll_read_time, nentry); } #endif exit: if (!recvbuf_is_contiguous) { if (NULL != receive_buf) { free (receive_buf); receive_buf = NULL; } } if (NULL != global_buf) { free (global_buf); global_buf = NULL; } if (NULL != sorted) { free (sorted); sorted = NULL; } if (NULL != global_iov_array) { free (global_iov_array); global_iov_array = NULL; } if (NULL != fview_count) { free (fview_count); fview_count = NULL; } if (NULL != decoded_iov) { free (decoded_iov); decoded_iov = NULL; } if (NULL != local_iov_array){ free(local_iov_array); local_iov_array=NULL; } if (NULL != displs) { free (displs); displs = NULL; } if (my_aggregator == fh->f_rank) { if (NULL != sorted_file_offsets){ free(sorted_file_offsets); sorted_file_offsets = NULL; } if (NULL != file_offsets_for_agg){ free(file_offsets_for_agg); file_offsets_for_agg = NULL; } if (NULL != memory_displacements){ free(memory_displacements); memory_displacements= NULL; } if (NULL != sendtype){ for (i = 0; i < fh->f_procs_per_group; i++) { if ( MPI_DATATYPE_NULL != sendtype[i] ) { ompi_datatype_destroy(&sendtype[i]); } } free(sendtype); sendtype=NULL; } if (NULL != disp_index){ free(disp_index); disp_index = NULL; } if ( NULL != blocklen_per_process){ for(l=0;l<fh->f_procs_per_group;l++){ if (NULL != blocklen_per_process[l]){ free(blocklen_per_process[l]); blocklen_per_process[l] = NULL; } } free(blocklen_per_process); blocklen_per_process = NULL; } if (NULL != displs_per_process){ for (l=0; i<fh->f_procs_per_group; l++){ if (NULL != displs_per_process[l]){ free(displs_per_process[l]); displs_per_process[l] = NULL; } } free(displs_per_process); displs_per_process = NULL; } if ( NULL != send_req ) { free ( send_req ); send_req = NULL; } } return ret; }
int mca_fcoll_two_phase_file_read_all (mca_io_ompio_file_t *fh, void *buf, int count, struct ompi_datatype_t *datatype, ompi_status_public_t *status) { int ret = OMPI_SUCCESS, i = 0, j = 0, interleave_count = 0, striping_unit = 0; MPI_Aint recv_buf_addr = 0; uint32_t iov_count = 0, ti = 0; struct iovec *decoded_iov = NULL, *temp_iov = NULL, *iov = NULL; size_t max_data = 0; long long_max_data = 0, long_total_bytes = 0; int domain_size=0, *count_my_req_per_proc=NULL, count_my_req_procs = 0; int count_other_req_procs; size_t *buf_indices=NULL; int *aggregator_list = NULL, local_count = 0, local_size = 0; int two_phase_num_io_procs=1; OMPI_MPI_OFFSET_TYPE start_offset = 0, end_offset = 0, fd_size = 0; OMPI_MPI_OFFSET_TYPE *start_offsets=NULL, *end_offsets=NULL; OMPI_MPI_OFFSET_TYPE *fd_start=NULL, *fd_end=NULL, min_st_offset = 0; Flatlist_node *flat_buf=NULL; mca_io_ompio_access_array_t *my_req=NULL, *others_req=NULL; #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN mca_common_ompio_print_entry nentry; #endif // if (opal_datatype_is_predefined(&datatype->super)) { // fh->f_flags = fh->f_flags | OMPIO_CONTIGUOUS_MEMORY; // } if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) { ret = fh->f_decode_datatype ((struct mca_io_ompio_file_t *)fh, datatype, count, buf, &max_data, &temp_iov, &iov_count); if (OMPI_SUCCESS != ret ){ goto exit; } recv_buf_addr = (size_t)(buf); decoded_iov = (struct iovec *) calloc (iov_count, sizeof(struct iovec)); for (ti = 0; ti < iov_count; ti++){ decoded_iov[ti].iov_base = (IOVBASE_TYPE *) ((OPAL_PTRDIFF_TYPE)temp_iov[ti].iov_base - recv_buf_addr); decoded_iov[ti].iov_len = temp_iov[ti].iov_len; #if DEBUG printf("d_offset[%d]: %ld, d_len[%d]: %ld\n", ti, (OPAL_PTRDIFF_TYPE)decoded_iov[ti].iov_base, ti, decoded_iov[ti].iov_len); #endif } } else{ max_data = count * datatype->super.size; } if ( MPI_STATUS_IGNORE != status ) { status->_ucount = max_data; } fh->f_get_num_aggregators (&two_phase_num_io_procs); if (-1 == two_phase_num_io_procs ){ ret = fh->f_set_aggregator_props ((struct mca_io_ompio_file_t *)fh, two_phase_num_io_procs, max_data); if (OMPI_SUCCESS != ret){ goto exit; } two_phase_num_io_procs = fh->f_final_num_aggrs; } if (two_phase_num_io_procs > fh->f_size){ two_phase_num_io_procs = fh->f_size; } aggregator_list = (int *) calloc (two_phase_num_io_procs, sizeof(int)); if (NULL == aggregator_list){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } for (i=0; i< two_phase_num_io_procs; i++){ aggregator_list[i] = i * fh->f_size / two_phase_num_io_procs; } ret = fh->f_generate_current_file_view ((struct mca_io_ompio_file_t *)fh, max_data, &iov, &local_count); if (OMPI_SUCCESS != ret){ goto exit; } long_max_data = (long) max_data; ret = fh->f_comm->c_coll.coll_allreduce (&long_max_data, &long_total_bytes, 1, MPI_LONG, MPI_SUM, fh->f_comm, fh->f_comm->c_coll.coll_allreduce_module); if ( OMPI_SUCCESS != ret ) { goto exit; } if (!(fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) { /* This datastructre translates between OMPIO->ROMIO its a little hacky!*/ /* But helps to re-use romio's code for handling non-contiguous file-type*/ /*Flattened datatype for ompio is in decoded_iov it translated into flatbuf*/ flat_buf = (Flatlist_node *)calloc(1, sizeof(Flatlist_node)); if ( NULL == flat_buf ){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } flat_buf->type = datatype; flat_buf->next = NULL; flat_buf->count = 0; flat_buf->indices = NULL; flat_buf->blocklens = NULL; if ( 0 < count ) { local_size = OMPIO_MAX(1,iov_count/count); } else { local_size = 0; } if ( 0 < local_size ) { flat_buf->indices = (OMPI_MPI_OFFSET_TYPE *)calloc(local_size, sizeof(OMPI_MPI_OFFSET_TYPE)); if (NULL == flat_buf->indices){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } flat_buf->blocklens = (OMPI_MPI_OFFSET_TYPE *)calloc(local_size, sizeof(OMPI_MPI_OFFSET_TYPE)); if ( NULL == flat_buf->blocklens ){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } } flat_buf->count = local_size; for (j = 0 ; j < local_size ; ++j) { flat_buf->indices[j] = (OMPI_MPI_OFFSET_TYPE)(intptr_t)decoded_iov[j].iov_base; flat_buf->blocklens[j] = decoded_iov[j].iov_len; } #if DEBUG printf("flat_buf count: %d\n", flat_buf->count); for(i=0;i<flat_buf->count;i++){ printf("%d: blocklen[%d] : %lld, indices[%d]: %lld\n", fh->f_rank, i, flat_buf->blocklens[i], i ,flat_buf->indices[i]); } #endif } #if DEBUG printf("%d: total_bytes:%ld, local_count: %d\n", fh->f_rank, long_total_bytes, local_count); for (i=0 ; i<local_count ; i++) { printf("%d: fcoll:two_phase:read_all:OFFSET:%ld,LENGTH:%ld\n", fh->f_rank, (size_t)iov[i].iov_base, (size_t)iov[i].iov_len); } #endif start_offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t)iov[0].iov_base; if ( 0 < local_count ) { end_offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t)iov[local_count-1].iov_base + (OMPI_MPI_OFFSET_TYPE)(intptr_t)iov[local_count-1].iov_len - 1; } else { end_offset = 0; } #if DEBUG printf("%d: START OFFSET:%ld, END OFFSET:%ld\n", fh->f_rank, (size_t)start_offset, (size_t)end_offset); #endif start_offsets = (OMPI_MPI_OFFSET_TYPE *)calloc (fh->f_size, sizeof(OMPI_MPI_OFFSET_TYPE)); if ( NULL == start_offsets ){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } end_offsets = (OMPI_MPI_OFFSET_TYPE *)calloc (fh->f_size, sizeof(OMPI_MPI_OFFSET_TYPE)); if (NULL == end_offsets){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } ret = fh->f_comm->c_coll.coll_allgather(&start_offset, 1, OMPI_OFFSET_DATATYPE, start_offsets, 1, OMPI_OFFSET_DATATYPE, fh->f_comm, fh->f_comm->c_coll.coll_allgather_module); if ( OMPI_SUCCESS != ret ){ goto exit; } ret = fh->f_comm->c_coll.coll_allgather(&end_offset, 1, OMPI_OFFSET_DATATYPE, end_offsets, 1, OMPI_OFFSET_DATATYPE, fh->f_comm, fh->f_comm->c_coll.coll_allgather_module); if ( OMPI_SUCCESS != ret ){ goto exit; } #if DEBUG for (i=0;i<fh->f_size;i++){ printf("%d: start[%d]:%ld,end[%d]:%ld\n", fh->f_rank,i, (size_t)start_offsets[i],i, (size_t)end_offsets[i]); } #endif for (i=1; i<fh->f_size; i++){ if ((start_offsets[i] < end_offsets[i-1]) && (start_offsets[i] <= end_offsets[i])){ interleave_count++; } } #if DEBUG printf("%d: interleave_count:%d\n", fh->f_rank,interleave_count); #endif ret = mca_fcoll_two_phase_domain_partition(fh, start_offsets, end_offsets, &min_st_offset, &fd_start, &fd_end, domain_size, &fd_size, striping_unit, two_phase_num_io_procs); if (OMPI_SUCCESS != ret){ goto exit; } #if DEBUG for (i=0;i<two_phase_num_io_procs;i++){ printf("fd_start[%d] : %lld, fd_end[%d] : %lld, local_count: %d\n", i, fd_start[i], i, fd_end[i], local_count); } #endif ret = mca_fcoll_two_phase_calc_my_requests (fh, iov, local_count, min_st_offset, fd_start, fd_end, fd_size, &count_my_req_procs, &count_my_req_per_proc, &my_req, &buf_indices, striping_unit, two_phase_num_io_procs, aggregator_list); if ( OMPI_SUCCESS != ret ){ goto exit; } ret = mca_fcoll_two_phase_calc_others_requests(fh, count_my_req_procs, count_my_req_per_proc, my_req, &count_other_req_procs, &others_req); if (OMPI_SUCCESS != ret ){ goto exit; } #if DEBUG printf("%d count_other_req_procs : %d\n", fh->f_rank, count_other_req_procs); #endif #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_rexch = MPI_Wtime(); #endif ret = two_phase_read_and_exch(fh, buf, datatype, others_req, iov, local_count, min_st_offset, fd_size, fd_start, fd_end, flat_buf, buf_indices, striping_unit, two_phase_num_io_procs, aggregator_list); if (OMPI_SUCCESS != ret){ goto exit; } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_rexch = MPI_Wtime(); read_exch += (end_rexch - start_rexch); nentry.time[0] = read_time; nentry.time[1] = rcomm_time; nentry.time[2] = read_exch; if (isread_aggregator(fh->f_rank, two_phase_num_io_procs, aggregator_list)){ nentry.aggregator = 1; } else{ nentry.aggregator = 0; } nentry.nprocs_for_coll = two_phase_num_io_procs; if (!mca_common_ompio_full_print_queue(fh->f_coll_read_time)){ mca_common_ompio_register_print_entry(fh->f_coll_read_time, nentry); } #endif exit: if (flat_buf != NULL){ if (flat_buf->blocklens != NULL){ free (flat_buf->blocklens); } if (flat_buf->indices != NULL){ free (flat_buf->indices); } free (flat_buf); } free (start_offsets); free (end_offsets); free (aggregator_list); free (fd_start); free (decoded_iov); free (buf_indices); free (count_my_req_per_proc); free (my_req); free (others_req); free (fd_end); return ret; }