int mca_fcoll_dynamic_gen2_file_read_all (mca_io_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 dynamic_gen2_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; 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_io_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 == (OPAL_PTRDIFF_TYPE) ftype_size) && opal_datatype_is_contiguous_memory_layout(&datatype->super,1) && 0 == lb ) { recvbuf_is_contiguous = true; } if (! recvbuf_is_contiguous ) { ret = fh->f_decode_datatype ((struct mca_io_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; } fh->f_get_num_aggregators ( &dynamic_gen2_num_io_procs); ret = fh->f_set_aggregator_props ((struct mca_io_ompio_file_t *) fh, dynamic_gen2_num_io_procs, max_data); if (OMPI_SUCCESS != ret){ goto exit; } my_aggregator = fh->f_procs_in_group[fh->f_aggregator_index]; /************************************************************************** ** 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 = fcoll_base_coll_allgather_array (&max_data, 1, MPI_LONG, total_bytes_per_process, 1, MPI_LONG, 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 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 mca_io_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 = fcoll_base_coll_allgather_array (&local_count, 1, MPI_INT, fview_count, 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 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 = fcoll_base_coll_allgatherv_array (local_iov_array, local_count, fh->f_iov_type, global_iov_array, fview_count, displs, fh->f_iov_type, 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 /**************************************************************************************** *** 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; } fh->f_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 *************************************************************/ fh->f_get_bytes_per_agg ( (int *) &bytes_per_cycle); 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] = (OPAL_PTRDIFF_TYPE)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] = (OPAL_PTRDIFF_TYPE)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] = (OPAL_PTRDIFF_TYPE)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] = (OPAL_PTRDIFF_TYPE) 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_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 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 ) { OPAL_PTRDIFF_TYPE mem_address; size_t remaining = 0; size_t temp_position = 0; remaining = bytes_received; while (remaining) { 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_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 = dynamic_gen2_num_io_procs; if (!fh->f_full_print_queue(READ_PRINT_QUEUE)){ fh->f_register_print_entry(READ_PRINT_QUEUE, 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; }
static ompi_datatype_t* __ompi_datatype_create_from_args( int32_t* i, MPI_Aint* a, ompi_datatype_t** d, int32_t type ) { ompi_datatype_t* datatype = NULL; switch(type){ /******************************************************************/ case MPI_COMBINER_DUP: /* should we duplicate d[0]? */ /* ompi_datatype_set_args( datatype, 0, NULL, 0, NULL, 1, d[0], MPI_COMBINER_DUP ); */ assert(0); /* shouldn't happen */ break; /******************************************************************/ case MPI_COMBINER_CONTIGUOUS: ompi_datatype_create_contiguous( i[0], d[0], &datatype ); ompi_datatype_set_args( datatype, 1, (const int **) &i, 0, NULL, 1, d, MPI_COMBINER_CONTIGUOUS ); break; /******************************************************************/ case MPI_COMBINER_VECTOR: ompi_datatype_create_vector( i[0], i[1], i[2], d[0], &datatype ); { const int* a_i[3] = {&i[0], &i[1], &i[2]}; ompi_datatype_set_args( datatype, 3, a_i, 0, NULL, 1, d, MPI_COMBINER_VECTOR ); } break; /******************************************************************/ case MPI_COMBINER_HVECTOR_INTEGER: case MPI_COMBINER_HVECTOR: ompi_datatype_create_hvector( i[0], i[1], a[0], d[0], &datatype ); { const int* a_i[2] = {&i[0], &i[1]}; ompi_datatype_set_args( datatype, 2, a_i, 1, a, 1, d, MPI_COMBINER_HVECTOR ); } break; /******************************************************************/ case MPI_COMBINER_INDEXED: /* TO CHECK */ ompi_datatype_create_indexed( i[0], &(i[1]), &(i[1+i[0]]), d[0], &datatype ); { const int* a_i[3] = {&i[0], &i[1], &(i[1+i[0]])}; ompi_datatype_set_args( datatype, 2 * i[0] + 1, a_i, 0, NULL, 1, d, MPI_COMBINER_INDEXED ); } break; /******************************************************************/ case MPI_COMBINER_HINDEXED_INTEGER: case MPI_COMBINER_HINDEXED: ompi_datatype_create_hindexed( i[0], &(i[1]), a, d[0], &datatype ); { const int* a_i[2] = {&i[0], &i[1]}; ompi_datatype_set_args( datatype, i[0] + 1, a_i, i[0], a, 1, d, MPI_COMBINER_HINDEXED ); } break; /******************************************************************/ case MPI_COMBINER_INDEXED_BLOCK: ompi_datatype_create_indexed_block( i[0], i[1], &(i[2]), d[0], &datatype ); { const int* a_i[3] = {&i[0], &i[1], &i[2]}; ompi_datatype_set_args( datatype, i[0] + 2, a_i, 0, NULL, 1, d, MPI_COMBINER_INDEXED_BLOCK ); } break; /******************************************************************/ case MPI_COMBINER_STRUCT_INTEGER: case MPI_COMBINER_STRUCT: ompi_datatype_create_struct( i[0], &(i[1]), a, d, &datatype ); { const int* a_i[2] = {&i[0], &i[1]}; ompi_datatype_set_args( datatype, i[0] + 1, a_i, i[0], a, i[0], d, MPI_COMBINER_STRUCT ); } break; /******************************************************************/ case MPI_COMBINER_SUBARRAY: ompi_datatype_create_subarray( i[0], &i[1 + 0 * i[0]], &i[1 + 1 * i[0]], &i[1 + 2 * i[0]], i[1 + 3 * i[0]], d[0], &datatype ); { const int* a_i[5] = {&i[0], &i[1 + 0 * i[0]], &i[1 + 1 * i[0]], &i[1 + 2 * i[0]], &i[1 + 3 * i[0]]}; ompi_datatype_set_args( datatype, 3 * i[0] + 2, a_i, 0, NULL, 1, d, MPI_COMBINER_SUBARRAY); } break; /******************************************************************/ case MPI_COMBINER_DARRAY: ompi_datatype_create_darray( i[0] /* size */, i[1] /* rank */, i[2] /* ndims */, &i[3 + 0 * i[0]], &i[3 + 1 * i[0]], &i[3 + 2 * i[0]], &i[3 + 3 * i[0]], i[3 + 4 * i[0]], d[0], &datatype ); { const int* a_i[8] = {&i[0], &i[1], &i[2], &i[3 + 0 * i[0]], &i[3 + 1 * i[0]], &i[3 + 2 * i[0]], &i[3 + 3 * i[0]], &i[3 + 4 * i[0]]}; ompi_datatype_set_args( datatype, 4 * i[0] + 4,a_i, 0, NULL, 1, d, MPI_COMBINER_DARRAY); } break; /******************************************************************/ case MPI_COMBINER_F90_REAL: case MPI_COMBINER_F90_COMPLEX: /*pArgs->i[0] = i[0][0]; pArgs->i[1] = i[1][0]; */ break; /******************************************************************/ case MPI_COMBINER_F90_INTEGER: /*pArgs->i[0] = i[0][0];*/ break; /******************************************************************/ case MPI_COMBINER_RESIZED: ompi_datatype_create_resized(d[0], a[0], a[1], &datatype); ompi_datatype_set_args( datatype, 0, NULL, 2, a, 1, d, MPI_COMBINER_RESIZED ); break; /******************************************************************/ case MPI_COMBINER_HINDEXED_BLOCK: ompi_datatype_create_hindexed_block( i[0], i[1], a, d[0], &datatype ); { const int* a_i[2] = {&i[0], &i[1]}; ompi_datatype_set_args( datatype, 2 + i[0], a_i, i[0], a, 1, d, MPI_COMBINER_HINDEXED_BLOCK ); } break; /******************************************************************/ default: break; } return datatype; }
int mca_fcoll_dynamic_file_write_all (mca_io_ompio_file_t *fh, void *buf, int count, struct ompi_datatype_t *datatype, ompi_status_public_t *status) { MPI_Aint total_bytes_written = 0; /* total bytes that have been written*/ MPI_Aint total_bytes = 0; /* total bytes to be written */ MPI_Aint bytes_to_write_in_cycle = 0; /* left to be written in a cycle*/ MPI_Aint bytes_per_cycle = 0; /* total written in each cycle by each process*/ int index = 0; 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 written from the current value from total_bytes_per_process */ int bytes_sent = 0, ret =0; int blocks=0, entries_per_aggregator=0; /* iovec structure and count of the buffer passed in */ uint32_t iov_count = 0; struct iovec *decoded_iov = NULL; int iov_index = 0; char *send_buf = NULL; size_t current_position = 0; struct iovec *local_iov_array=NULL, *global_iov_array=NULL; local_io_array *file_offsets_for_agg=NULL; /* global iovec at the writers that contain the iovecs created from file_set_view */ uint32_t total_fview_count = 0; int local_count = 0, temp_pindex; int *fview_count = NULL, *disp_index=NULL, *temp_disp_index=NULL; int current_index = 0, temp_index=0; char *global_buf = NULL; MPI_Aint global_count = 0; /* array that contains the sorted indices of the global_iov */ int *sorted = NULL, *sorted_file_offsets=NULL; int *displs = NULL; size_t max_data = 0; int **blocklen_per_process=NULL; MPI_Aint **displs_per_process=NULL, *memory_displacements=NULL; ompi_datatype_t **recvtype = NULL; MPI_Aint *total_bytes_per_process = NULL; MPI_Request *send_req=NULL, *recv_req=NULL; int datatype_size, recv_req_count=0; #if TIME_BREAKDOWN double start_time=0.0, end_time=0.0, start_time2=0.0, end_time2=0.0; double total=0.0 , total_io=0.0, max_io=0.0; /* max_pp=0.0;*/ double start_ptime=0.0, end_ptime=0.0, tpw=0.0; /* max_tpw=0.0;*/ double start_cio_array=0.0, end_cio_array=0.0, tcio_array=0.0;/* max_cio=0.0;*/ double start_sr=0.0, end_sr=0.0, tsr=0.0;/* max_sr=0.0;*/ double comm_time = 0.0, max_comm_time=0.0; double write_time = 0.0, max_write_time=0.0; #endif #if TIME_BREAKDOWN start_time = MPI_Wtime(); #endif if (opal_datatype_is_contiguous_memory_layout(&datatype->super,1)) { fh->f_flags |= OMPIO_CONTIGUOUS_MEMORY; } /************************************************************************** ** In case the data is not contigous in memory, decode it into an iovec ** **************************************************************************/ if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) { ret = ompi_io_ompio_decode_datatype (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; } if (! (fh->f_flags & OMPIO_AGGREGATOR_IS_SET)) { ret = ompi_io_ompio_set_aggregator_props (fh, mca_fcoll_dynamic_num_io_procs, max_data); if (OMPI_SUCCESS != ret){ goto exit; } } if (-1 == mca_fcoll_dynamic_num_io_procs) { mca_fcoll_dynamic_num_io_procs = 1; } #if TIME_BREAKDOWN if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { start_time = MPI_Wtime(); } #endif 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; } ret = ompi_io_ompio_allgather_array (&max_data, 1, MPI_LONG, total_bytes_per_process, 1, MPI_LONG, fh->f_aggregator_index, fh->f_procs_in_group, fh->f_procs_per_group, fh->f_comm); if( OMPI_SUCCESS != ret){ goto exit; } 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; } /********************************************************************* *** Generate the File offsets/lengths corresponding to this write *** ********************************************************************/ ret = ompi_io_ompio_generate_current_file_view(fh, max_data, &local_iov_array, &local_count); if (ret != OMPI_SUCCESS){ goto exit; } #if DEBUG_ON for (i=0 ; i<local_count ; i++) { printf("Local offset-length pair for rank:%d \n", fh->f_rank); printf("%d: OFFSET: %p LENGTH: %lld\n", fh->f_rank, iov[i].iov_base, iov[i].iov_len); } #endif /************************************************************* *** 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; } ret = ompi_io_ompio_allgather_array (&local_count, 1, MPI_INT, fview_count, 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; } 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 (fh->f_procs_in_group[fh->f_aggregator_index] == 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 (fh->f_flags & OMPIO_UNIFORM_FVIEW) { ret = ompi_io_ompio_allgather_array (local_iov_array, local_count, fh->f_iov_type, global_iov_array, local_count, fh->f_iov_type, fh->f_aggregator_index, fh->f_procs_in_group, fh->f_procs_per_group, fh->f_comm); if (OMPI_SUCCESS != ret){ goto exit; } } else { ret = ompi_io_ompio_allgatherv_array (local_iov_array, local_count, fh->f_iov_type, global_iov_array, fview_count, displs, fh->f_iov_type, fh->f_aggregator_index, fh->f_procs_in_group, fh->f_procs_per_group, fh->f_comm); if (OMPI_SUCCESS != ret){ goto exit; } } /* sort it */ 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_io_ompio_sort_iovec (global_iov_array, total_fview_count, sorted); } if (NULL != local_iov_array){ free(local_iov_array); local_iov_array = NULL; } if (NULL != displs){ free(displs); displs=NULL; } #if DEBUG_ON if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { uint32_t tv=0; for (tv=0 ; tv<total_fview_count ; tv++) { printf("%d: OFFSET: %lld LENGTH: %ld\n", fh->f_rank, global_iov_array[sorted[tv]].offset, global_iov_array[sorted[tv]].length); } } #endif if (fh->f_procs_in_group[fh->f_aggregator_index] == 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; } } bytes_per_cycle = mca_fcoll_dynamic_cycle_buffer_size; cycles = ceil((double)total_bytes/bytes_per_cycle); n = 0; bytes_remaining = 0; current_index = 0; #if TIME_BREAKDOWN end_time = MPI_Wtime(); total = end_time-start_time; start_time2 = MPI_Wtime(); #endif for (index = 0; index < cycles; index++) { /* Getting ready for next cycle Initializing and freeing buffers*/ if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { if (NULL == recvtype){ recvtype = (ompi_datatype_t **) malloc (fh->f_procs_per_group * sizeof(ompi_datatype_t *)); if (NULL == recvtype) { 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++){ 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; } } if (cycles-1 == index) { bytes_to_write_in_cycle = total_bytes - bytes_per_cycle*index; } else { bytes_to_write_in_cycle = bytes_per_cycle; } #if DEBUG_ON if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { printf ("****%d: CYCLE %d Bytes %lld**********\n", fh->f_rank, index, bytes_to_write_in_cycle); } #endif /********************************************************** **Gather the Data from all the processes at the writers ** *********************************************************/ /* Calculate how much data will be contributed in this cycle by each process*/ bytes_sent = 0; #if DEBUG_ON printf("bytes_to_write_in_cycle: %ld, cycle : %d\n", bytes_to_write_in_cycle, index); #endif /* The blocklen and displs calculation only done at aggregators!*/ #if TIME_BREAKDOWN start_cio_array = MPI_Wtime(); #endif while (bytes_to_write_in_cycle) { 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) { if (bytes_remaining <= bytes_to_write_in_cycle) { if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { blocklen_per_process[n][disp_index[n] - 1] = bytes_remaining; displs_per_process[n][disp_index[n] - 1] = (OPAL_PTRDIFF_TYPE)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_sent += bytes_remaining; } current_index ++; bytes_to_write_in_cycle -= bytes_remaining; bytes_remaining = 0; if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { /* In this cases the length is consumed so allocating for next displacement and blocklength*/ 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; } continue; } else { if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { blocklen_per_process[n][disp_index[n] - 1] = bytes_to_write_in_cycle; displs_per_process[n][disp_index[n] - 1] = (OPAL_PTRDIFF_TYPE)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_sent += bytes_to_write_in_cycle; } bytes_remaining -= bytes_to_write_in_cycle; bytes_to_write_in_cycle = 0; break; } } else { if (bytes_to_write_in_cycle < (MPI_Aint) global_iov_array[sorted[current_index]].iov_len) { if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { blocklen_per_process[n][disp_index[n] - 1] = bytes_to_write_in_cycle; displs_per_process[n][disp_index[n] - 1] = (OPAL_PTRDIFF_TYPE)global_iov_array[sorted[current_index]].iov_base ; } if (fh->f_procs_in_group[n] == fh->f_rank) { bytes_sent += bytes_to_write_in_cycle; } bytes_remaining = global_iov_array[sorted[current_index]].iov_len - bytes_to_write_in_cycle; bytes_to_write_in_cycle = 0; break; } else { if (fh->f_procs_in_group[fh->f_aggregator_index] == 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] = (OPAL_PTRDIFF_TYPE) 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; /*realloc for next blocklength and assign this displacement and check for next displs as the total length of this entry has been consumed!*/ } if (fh->f_procs_in_group[n] == fh->f_rank) { bytes_sent += global_iov_array[sorted[current_index]].iov_len; } bytes_to_write_in_cycle -= global_iov_array[sorted[current_index]].iov_len; current_index ++; continue; } } } #if TIME_BREAKDOWN start_sr = MPI_Wtime(); #endif /* Calculate the displacement on where to put the data and allocate the recieve buffer (global_buf) */ if (fh->f_procs_in_group[fh->f_aggregator_index] == 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 DEBUG_ON printf("%d: cycle: %d, bytes_sent: %d\n ",fh->f_rank,index, bytes_sent); printf("%d : Entries per aggregator : %d\n",fh->f_rank,entries_per_aggregator); #endif if (entries_per_aggregator > 0){ file_offsets_for_agg = (local_io_array *) malloc(entries_per_aggregator*sizeof(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; 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]; file_offsets_for_agg[temp_index].process_id = i; file_offsets_for_agg[temp_index].offset = displs_per_process[i][j]; temp_index++; #if DEBUG_ON printf("************Cycle: %d, Aggregator: %d ***************\n", index+1,fh->f_rank); 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 } } } } else{ continue; } /* Sort the displacements for each aggregator*/ local_heap_sort (file_offsets_for_agg, entries_per_aggregator, sorted_file_offsets); /*create contiguous memory displacements based on blocklens on the same displs array and map it to this aggregator's actual file-displacements (this is in the io-array created above)*/ 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; } 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"); return OMPI_ERR_OUT_OF_RESOURCE; } /*Now update the displacements array with memory offsets*/ global_count = 0; for (i=0;i<entries_per_aggregator;i++){ temp_pindex = file_offsets_for_agg[sorted_file_offsets[i]].process_id; displs_per_process[temp_pindex][temp_disp_index[temp_pindex]] = memory_displacements[sorted_file_offsets[i]]; if (temp_disp_index[temp_pindex] < disp_index[temp_pindex]) temp_disp_index[temp_pindex] += 1; else{ printf("temp_disp_index[%d]: %d is greater than disp_index[%d]: %d\n", temp_pindex, temp_disp_index[temp_pindex], temp_pindex, disp_index[temp_pindex]); } global_count += file_offsets_for_agg[sorted_file_offsets[i]].length; } if (NULL != temp_disp_index){ free(temp_disp_index); temp_disp_index = NULL; } #if DEBUG_ON printf("************Cycle: %d, Aggregator: %d ***************\n", index+1,fh->f_rank); for (i=0;i<fh->f_procs_per_group; i++){ for(j=0;j<disp_index[i];j++){ if (blocklen[i][j] > 0){ 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); } } } 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\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]]); } printf("%d : global_count : %ld, bytes_sent : %d\n", fh->f_rank,global_count, bytes_sent); #endif global_buf = (char *) malloc (global_count); if (NULL == global_buf){ opal_output(1, "OUT OF MEMORY"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } recv_req_count = 0; for (i=0;i<fh->f_procs_per_group; i++){ ompi_datatype_create_hindexed(disp_index[i], blocklen_per_process[i], displs_per_process[i], MPI_BYTE, &recvtype[i]); ompi_datatype_commit(&recvtype[i]); MPI_Type_size (recvtype[i], &datatype_size); if (datatype_size){ recv_req = (MPI_Request *)realloc ((void *)recv_req, (recv_req_count + 1)*sizeof(MPI_Request)); ret = MCA_PML_CALL(irecv(global_buf, 1, recvtype[i], fh->f_procs_in_group[i], 123, fh->f_comm, &recv_req[recv_req_count])); recv_req_count++; if (OMPI_SUCCESS != ret){ goto exit; } } } } #if TIME_BREAKDOWN end_cio_array = MPI_Wtime(); tcio_array = end_cio_array - start_cio_array; #endif if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) { send_buf = &((char*)buf)[total_bytes_written]; } else if (bytes_sent) { /* allocate a send buffer and copy the data that needs to be sent into it in case the data is non-contigous in memory */ OPAL_PTRDIFF_TYPE mem_address; size_t remaining = 0; size_t temp_position = 0; send_buf = malloc (bytes_sent); if (NULL == send_buf) { opal_output (1, "OUT OF MEMORY\n"); return OMPI_ERR_OUT_OF_RESOURCE; } remaining = bytes_sent; while (remaining) { mem_address = (OPAL_PTRDIFF_TYPE) (decoded_iov[iov_index].iov_base) + current_position; if (remaining >= (decoded_iov[iov_index].iov_len - current_position)) { memcpy (send_buf+temp_position, (IOVBASE_TYPE *)mem_address, 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 (send_buf+temp_position, (IOVBASE_TYPE *) mem_address, remaining); current_position = current_position + remaining; remaining = 0; } } } total_bytes_written += bytes_sent; /* Gather the sendbuf from each process in appropritate locations in aggregators*/ send_req = (MPI_Request *) malloc (sizeof(MPI_Request)); if (NULL == send_req){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } if (bytes_sent){ ret = MCA_PML_CALL(isend(send_buf, bytes_sent, MPI_BYTE, fh->f_procs_in_group[fh->f_aggregator_index], 123, MCA_PML_BASE_SEND_STANDARD, fh->f_comm, send_req)); if ( OMPI_SUCCESS != ret ){ goto exit; } ret = ompi_request_wait(send_req, MPI_STATUS_IGNORE); if (OMPI_SUCCESS != ret){ goto exit; } } if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { ret = ompi_request_wait_all (recv_req_count, recv_req, MPI_STATUS_IGNORE); if (OMPI_SUCCESS != ret){ goto exit; } } #if DEBUG_ON if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank){ printf("************Cycle: %d, Aggregator: %d ***************\n", index+1,fh->f_rank); for (i=0 ; i<global_count/4 ; i++) printf (" RECV %d \n",((int *)global_buf)[i]); } #endif #if TIME_BREAKDOWN end_sr = MPI_Wtime(); tsr = end_sr - start_sr; comm_time += tsr; #endif if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) { if (NULL != send_buf) { free (send_buf); send_buf = NULL; } } /********************************************************** **************** DONE GATHERING OF DATA ****************** *********************************************************/ /********************************************************** ******* Create the io array, and pass it to fbtl ********* *********************************************************/ if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { 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"); return OMPI_ERR_OUT_OF_RESOURCE; } fh->f_num_of_io_entries = 0; /*First entry for every aggregator*/ fh->f_io_array[fh->f_num_of_io_entries].offset = (IOVBASE_TYPE *)file_offsets_for_agg[sorted_file_offsets[0]].offset; fh->f_io_array[fh->f_num_of_io_entries].length = file_offsets_for_agg[sorted_file_offsets[0]].length; fh->f_io_array[fh->f_num_of_io_entries].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 the enrties are contiguous merge them, else make a new entry */ 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 *)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 TIME_BREAKDOWN start_ptime = MPI_Wtime(); #endif if (fh->f_num_of_io_entries) { if (OMPI_SUCCESS != fh->f_fbtl->fbtl_pwritev (fh, NULL)) { opal_output (1, "WRITE FAILED\n"); ret = OMPI_ERROR; goto exit; } } #if TIME_BREAKDOWN end_ptime = MPI_Wtime(); tpw = end_ptime - start_ptime; write_time += tpw; #endif } if (NULL != send_req){ free(send_req); send_req = NULL; } if (fh->f_procs_in_group[fh->f_aggregator_index] == 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 != recvtype){ free(recvtype); recvtype=NULL; } if (NULL != recv_req){ free(recv_req); recv_req = NULL; } if (NULL != global_buf) { free (global_buf); global_buf = NULL; } } } #if TIME_BREAKDOWN end_time2 = MPI_Wtime(); total_io = end_time2-start_time2; fh->f_comm->c_coll.coll_allreduce (&total_io, &max_io, 1, MPI_DOUBLE, MPI_MAX, fh->f_comm, fh->f_comm->c_coll.coll_allreduce_module); fh->f_comm->c_coll.coll_allreduce (&comm_time, &max_comm_time, 1, MPI_DOUBLE, MPI_SUM, fh->f_comm, fh->f_comm->c_coll.coll_allreduce_module); fh->f_comm->c_coll.coll_allreduce (&write_time, &max_write_time, 1, MPI_DOUBLE, MPI_SUM, fh->f_comm, fh->f_comm->c_coll.coll_allreduce_module); if (0 == fh->f_rank){ printf ("Max Exchange and write ---- %f\n", max_io); printf ("AVG pwrite time : %f \n", max_write_time/mca_fcoll_dynamic_num_io_procs); printf ("AVG communication time : %f\n", max_comm_time/fh->f_size); } fh->f_comm->c_coll.coll_allreduce (&comm_time, &max_comm_time, 1, MPI_DOUBLE, MPI_MAX, fh->f_comm, fh->f_comm->c_coll.coll_allreduce_module); fh->f_comm->c_coll.coll_allreduce (&write_time, &max_write_time, 1, MPI_DOUBLE, MPI_MAX, fh->f_comm, fh->f_comm->c_coll.coll_allreduce_module); if (0 == fh->f_rank){ printf ("MAX pwrite time : %f \n", max_write_time); printf ("MAX communication time : %f\n", max_comm_time); } fh->f_comm->c_coll.coll_allreduce (&comm_time, &max_comm_time, 1, MPI_DOUBLE, MPI_MIN, fh->f_comm, fh->f_comm->c_coll.coll_allreduce_module); if (0 == fh->f_rank){ printf ("MIN communication time : %f\n", max_comm_time); } #endif exit : if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { if (NULL != fh->f_io_array) { free (fh->f_io_array); fh->f_io_array = NULL; } if (NULL != disp_index){ free(disp_index); disp_index = NULL; } if (NULL != recvtype){ free(recvtype); recvtype=NULL; } if (NULL != recv_req){ free(recv_req); recv_req = NULL; } if (NULL != global_buf) { free (global_buf); global_buf = NULL; } 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; } if (NULL != displs_per_process[l]){ free(displs_per_process[l]); displs_per_process[l] = NULL; } } if (NULL != blocklen_per_process){ free(blocklen_per_process); blocklen_per_process = NULL; } if (NULL != displs_per_process){ free(displs_per_process); displs_per_process = 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 != send_req){ free(send_req); send_req = NULL; } return OMPI_SUCCESS; }
int mca_fcoll_dynamic_file_read_all (mca_io_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; 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 dynamic_num_io_procs; size_t max_data = 0; int *bytes_per_process = NULL; MPI_Aint *total_bytes_per_process = NULL; ompi_datatype_t **sendtype = NULL; MPI_Request *send_req=NULL, *recv_req=NULL; #if 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; print_entry nentry; #endif // if (opal_datatype_is_contiguous_memory_layout(&datatype->super,1)) { // fh->f_flags |= OMPIO_CONTIGUOUS_MEMORY; // } /************************************************************************** ** In case the data is not contigous in memory, decode it into an iovec ** **************************************************************************/ if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) { ret = fh->f_decode_datatype ((struct mca_io_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; } fh->f_get_num_aggregators ( &dynamic_num_io_procs); ret = fh->f_set_aggregator_props ((struct mca_io_ompio_file_t *) fh, dynamic_num_io_procs, max_data); if (OMPI_SUCCESS != ret){ goto exit; } 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; } ret = fh->f_allgather_array (&max_data, 1, MPI_LONG, total_bytes_per_process, 1, MPI_LONG, fh->f_aggregator_index, fh->f_procs_in_group, fh->f_procs_per_group, fh->f_comm); if (OMPI_SUCCESS != ret){ goto exit; } 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; } /********************************************************************* *** Generate the File offsets/lengths corresponding to this write *** ********************************************************************/ ret = fh->f_generate_current_file_view ((struct mca_io_ompio_file_t *) fh, max_data, &local_iov_array, &local_count); if (ret != OMPI_SUCCESS){ goto exit; } /* #########################################################*/ /************************************************************* *** 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; } ret = fh->f_allgather_array (&local_count, 1, MPI_INT, fview_count, 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; } 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 (fh->f_procs_in_group[fh->f_aggregator_index] == 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; } } ret = fh->f_allgatherv_array (local_iov_array, local_count, fh->f_iov_type, global_iov_array, fview_count, displs, fh->f_iov_type, fh->f_aggregator_index, fh->f_procs_in_group, fh->f_procs_per_group, fh->f_comm); if (OMPI_SUCCESS != ret){ goto exit; } /* sort it */ 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; } fh->f_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 (fh->f_procs_in_group[fh->f_aggregator_index] == 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 if (fh->f_procs_in_group[fh->f_aggregator_index] == 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; } } /* * Calculate how many bytes are read in each cycle */ fh->f_get_bytes_per_agg ( (int *) &bytes_per_cycle); cycles = ceil((double)total_bytes/bytes_per_cycle); n = 0; bytes_remaining = 0; current_index = 0; #if TIME_BREAKDOWN start_rexch = MPI_Wtime(); #endif for (index = 0; index < cycles; index++) { /* Getting ready for next cycle Initializing and freeing buffers */ if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { if (NULL == sendtype){ 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++){ 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; } } 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 (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { printf ("****%d: CYCLE %d Bytes %d**********\n", fh->f_rank, index, bytes_to_write_in_cycle); } #endif /* Calculate how much data will be contributed in this cycle by each process*/ bytes_received = 0; while (bytes_to_read_in_cycle) { 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) { if (bytes_remaining <= bytes_to_read_in_cycle) { if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { blocklen_per_process[n][disp_index[n] - 1] = bytes_remaining; displs_per_process[n][disp_index[n] - 1] = (OPAL_PTRDIFF_TYPE)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_remaining; } current_index ++; bytes_to_read_in_cycle -= bytes_remaining; bytes_remaining = 0; if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { 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; } continue; } else { if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { blocklen_per_process[n][disp_index[n] - 1] = bytes_to_read_in_cycle; displs_per_process[n][disp_index[n] - 1] = (OPAL_PTRDIFF_TYPE)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 { if (bytes_to_read_in_cycle < (MPI_Aint) global_iov_array[sorted[current_index]].iov_len) { if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { blocklen_per_process[n][disp_index[n] - 1] = bytes_to_read_in_cycle; displs_per_process[n][disp_index[n] - 1] = (OPAL_PTRDIFF_TYPE)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 { if (fh->f_procs_in_group[fh->f_aggregator_index] == 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] = (OPAL_PTRDIFF_TYPE) 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; } } } /* Calculate the displacement on where to put the data and allocate the recieve buffer (global_buf) */ if (fh->f_procs_in_group[fh->f_aggregator_index] == 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 = (local_io_array *) malloc(entries_per_aggregator*sizeof(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; } 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; } 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; } 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[fh->f_num_of_io_entries].offset = (IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[0]].offset; fh->f_io_array[fh->f_num_of_io_entries].length = file_offsets_for_agg[sorted_file_offsets[0]].length; fh->f_io_array[fh->f_num_of_io_entries].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 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 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; } 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; } #if TIME_BREAKDOWN start_rcomm_time = MPI_Wtime(); #endif for (i=0;i<fh->f_procs_per_group;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 TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif } /********************************************************** ********* Scatter the Data from the readers ************** *********************************************************/ if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) { 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 TIME_BREAKDOWN start_rcomm_time = MPI_Wtime(); #endif recv_req = (MPI_Request *) malloc (sizeof (MPI_Request)); if (NULL == recv_req){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } ret = MCA_PML_CALL(irecv(receive_buf, bytes_received, MPI_BYTE, fh->f_procs_in_group[fh->f_aggregator_index], 123, fh->f_comm, recv_req)); if (OMPI_SUCCESS != ret){ goto exit; } if (fh->f_procs_in_group[fh->f_aggregator_index] == 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 (!(fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) { OPAL_PTRDIFF_TYPE mem_address; size_t remaining = 0; size_t temp_position = 0; remaining = bytes_received; while (remaining) { 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 TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += end_rcomm_time - start_rcomm_time; #endif if (NULL != recv_req){ free(recv_req); recv_req = NULL; } if (fh->f_procs_in_group[fh->f_aggregator_index] == 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; } for (i = 0; i < fh->f_procs_per_group; i++) ompi_datatype_destroy(sendtype+i); if (NULL != sendtype){ free(sendtype); sendtype=NULL; } if (NULL != send_req){ free(send_req); send_req = 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 != bytes_per_process){ free(bytes_per_process); bytes_per_process =NULL; } if (NULL != memory_displacements){ free(memory_displacements); memory_displacements= NULL; } } } #if 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 (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) nentry.aggregator = 1; else nentry.aggregator = 0; nentry.nprocs_for_coll = dynamic_num_io_procs; if (!fh->f_full_print_queue(READ_PRINT_QUEUE)){ fh->f_register_print_entry(READ_PRINT_QUEUE, nentry); } #endif exit: 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 (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { 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; } } return ret; }
static int two_phase_exchage_data(mca_io_ompio_file_t *fh, void *buf, char *write_buf, struct iovec *offset_length, int *send_size,int *start_pos, int *recv_size, OMPI_MPI_OFFSET_TYPE off, OMPI_MPI_OFFSET_TYPE size, int *count, int *partial_recv, int *sent_to_proc, int contig_access_count, OMPI_MPI_OFFSET_TYPE min_st_offset, OMPI_MPI_OFFSET_TYPE fd_size, OMPI_MPI_OFFSET_TYPE *fd_start, OMPI_MPI_OFFSET_TYPE *fd_end, Flatlist_node *flat_buf, mca_io_ompio_access_array_t *others_req, int *send_buf_idx, int *curr_to_proc, int *done_to_proc, int iter, size_t *buf_idx,MPI_Aint buftype_extent, int striping_unit, int *aggregator_list, int *hole){ int *tmp_len=NULL, sum, *srt_len=NULL, nprocs_recv, nprocs_send, k,i,j; int ret=OMPI_SUCCESS; MPI_Request *requests=NULL, *send_req=NULL; ompi_datatype_t **recv_types=NULL; OMPI_MPI_OFFSET_TYPE *srt_off=NULL; char **send_buf = NULL; #if TIME_BREAKDOWN start_comm_time = MPI_Wtime(); #endif ret = fh->f_comm->c_coll.coll_alltoall (recv_size, 1, MPI_INT, send_size, 1, MPI_INT, fh->f_comm, fh->f_comm->c_coll.coll_alltoall_module); if ( OMPI_SUCCESS != ret ){ return ret; } nprocs_recv = 0; for (i=0;i<fh->f_size;i++){ if (recv_size[i]){ nprocs_recv++; } } recv_types = (ompi_datatype_t **) malloc (( nprocs_recv + 1 ) * sizeof(ompi_datatype_t *)); if ( NULL == recv_types ){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } tmp_len = (int *) malloc(fh->f_size*sizeof(int)); if ( NULL == tmp_len ) { return OMPI_ERR_OUT_OF_RESOURCE; } j = 0; for (i=0;i<fh->f_size;i++){ if (recv_size[i]) { if (partial_recv[i]) { k = start_pos[i] + count[i] - 1; tmp_len[i] = others_req[i].lens[k]; others_req[i].lens[k] = partial_recv[i]; } ompi_datatype_create_hindexed(count[i], &(others_req[i].lens[start_pos[i]]), &(others_req[i].mem_ptrs[start_pos[i]]), MPI_BYTE, recv_types+j); ompi_datatype_commit(recv_types+j); j++; } } sum = 0; for (i=0;i<fh->f_size;i++) sum += count[i]; srt_off = (OMPI_MPI_OFFSET_TYPE *) malloc((sum+1)*sizeof(OMPI_MPI_OFFSET_TYPE)); if ( NULL == srt_off ){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } srt_len = (int *) malloc((sum+1)*sizeof(int)); if ( NULL == srt_len ) { ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } two_phase_heap_merge(others_req, count, srt_off, srt_len, start_pos, fh->f_size,fh->f_rank, nprocs_recv, sum); for (i=0; i<fh->f_size; i++) if (partial_recv[i]) { k = start_pos[i] + count[i] - 1; others_req[i].lens[k] = tmp_len[i]; } if ( NULL != tmp_len ){ free(tmp_len); } *hole = 0; if (off != srt_off[0]){ *hole = 1; } else{ for (i=1;i<sum;i++){ if (srt_off[i] <= srt_off[0] + srt_len[0]){ int new_len = srt_off[i] + srt_len[i] - srt_off[0]; if(new_len > srt_len[0]) srt_len[0] = new_len; } else break; } if (i < sum || size != srt_len[0]) *hole = 1; } if ( NULL != srt_off ){ free(srt_off); } if ( NULL != srt_len ){ free(srt_len); } if (nprocs_recv){ if (*hole){ if (off > 0){ fh->f_io_array = (mca_io_ompio_io_array_t *)malloc (sizeof(mca_io_ompio_io_array_t)); if (NULL == fh->f_io_array) { opal_output(1, "OUT OF MEMORY\n"); return OMPI_ERR_OUT_OF_RESOURCE; } fh->f_io_array[0].offset =(IOVBASE_TYPE *)(intptr_t) off; fh->f_num_of_io_entries = 1; fh->f_io_array[0].length = size; fh->f_io_array[0].memory_address = write_buf; if (fh->f_num_of_io_entries){ if (OMPI_SUCCESS != fh->f_fbtl->fbtl_preadv (fh, NULL)) { opal_output(1, "READ FAILED\n"); return OMPI_ERROR; } } } fh->f_num_of_io_entries = 0; if (NULL != fh->f_io_array) { free (fh->f_io_array); fh->f_io_array = NULL; } } } nprocs_send = 0; for (i=0; i <fh->f_size; i++) if (send_size[i]) nprocs_send++; #if DEBUG_ON printf("%d : nprocs_send : %d\n", fh->f_rank,nprocs_send); #endif requests = (MPI_Request *) malloc((nprocs_send+nprocs_recv+1)*sizeof(MPI_Request)); if ( NULL == requests ){ return OMPI_ERR_OUT_OF_RESOURCE; } j = 0; for (i=0; i<fh->f_size; i++) { if (recv_size[i]) { ret = MCA_PML_CALL(irecv(MPI_BOTTOM, 1, recv_types[j], i, fh->f_rank+i+100*iter, fh->f_comm, requests+j)); if ( OMPI_SUCCESS != ret ){ goto exit; } j++; } } send_req = requests + nprocs_recv; if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) { j = 0; for (i=0; i <fh->f_size; i++) if (send_size[i]) { ret = MCA_PML_CALL(isend(((char *) buf) + buf_idx[i], send_size[i], MPI_BYTE, i, fh->f_rank+i+100*iter, MCA_PML_BASE_SEND_STANDARD, fh->f_comm, send_req+j)); if ( OMPI_SUCCESS != ret ){ goto exit; } j++; buf_idx[i] += send_size[i]; } } else if(nprocs_send && (!(fh->f_flags & OMPIO_CONTIGUOUS_MEMORY))){ send_buf = (char **) malloc(fh->f_size*sizeof(char*)); if ( NULL == send_buf ){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } for (i=0; i < fh->f_size; i++){ if (send_size[i]) { send_buf[i] = (char *) malloc(send_size[i]); if ( NULL == send_buf[i] ){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } } } ret = two_phase_fill_send_buffer(fh, buf,flat_buf, send_buf, offset_length, send_size, send_req,sent_to_proc, contig_access_count, min_st_offset, fd_size, fd_start, fd_end, send_buf_idx, curr_to_proc, done_to_proc, iter, buftype_extent, striping_unit, aggregator_list); if ( OMPI_SUCCESS != ret ){ goto exit; } } for (i=0; i<nprocs_recv; i++) ompi_datatype_destroy(recv_types+i); if (NULL != recv_types){ free(recv_types); recv_types = NULL; } ret = ompi_request_wait_all (nprocs_send+nprocs_recv, requests, MPI_STATUS_IGNORE); if ( NULL != requests ){ free(requests); } #if TIME_BREAKDOWN end_comm_time = MPI_Wtime(); comm_time += (end_comm_time - start_comm_time); #endif exit: return ret; }
static int two_phase_exchange_data(mca_io_ompio_file_t *fh, void *buf, struct iovec *offset_len, int *send_size, int *start_pos, int *recv_size, int *count, int *partial_send, int *recd_from_proc, int contig_access_count, OMPI_MPI_OFFSET_TYPE min_st_offset, OMPI_MPI_OFFSET_TYPE fd_size, OMPI_MPI_OFFSET_TYPE *fd_start, OMPI_MPI_OFFSET_TYPE *fd_end, Flatlist_node *flat_buf, mca_io_ompio_access_array_t *others_req, int iter, size_t *buf_idx, MPI_Aint buftype_extent, int striping_unit, int *aggregator_list) { int i=0, j=0, k=0, tmp=0, nprocs_recv=0, nprocs_send=0; int ret = OMPI_SUCCESS; char **recv_buf = NULL; MPI_Request *requests=NULL; MPI_Datatype send_type; #if TIME_BREAKDOWN start_rcomm_time = MPI_Wtime(); #endif ret = fh->f_comm->c_coll.coll_alltoall (send_size, 1, MPI_INT, recv_size, 1, MPI_INT, fh->f_comm, fh->f_comm->c_coll.coll_alltoall_module); if ( OMPI_SUCCESS != ret ){ goto exit; } #if DEBUG for (i=0; i<fh->f_size; i++){ printf("%d: RS[%d]: %d\n", fh->f_rank, i, recv_size[i]); } #endif nprocs_recv = 0; for (i=0; i < fh->f_size; i++) if (recv_size[i]) nprocs_recv++; nprocs_send = 0; for (i=0; i< fh->f_size; i++) if (send_size[i]) nprocs_send++; requests = (MPI_Request *) malloc((nprocs_send+nprocs_recv+1) * sizeof(MPI_Request)); if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) { j = 0; for (i=0; i < fh->f_size; i++){ if (recv_size[i]){ ret = MCA_PML_CALL(irecv(((char *) buf)+ buf_idx[i], recv_size[i], MPI_BYTE, i, fh->f_rank+i+100*iter, fh->f_comm, requests+j)); if ( OMPI_SUCCESS != ret ){ return ret; } j++; buf_idx[i] += recv_size[i]; } } } else{ recv_buf = (char **)malloc(fh->f_size * sizeof(char *)); if (NULL == recv_buf){ ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } for (i=0; i < fh->f_size; i++) if(recv_size[i]) recv_buf[i] = (char *) malloc (recv_size[i] * sizeof(char)); j = 0; for(i=0; i<fh->f_size; i++) if (recv_size[i]) { ret = MCA_PML_CALL(irecv(recv_buf[i], recv_size[i], MPI_BYTE, i, fh->f_rank+i+100*iter, fh->f_comm, requests+j)); j++; } } j = 0; for (i = 0; i< fh->f_size; i++){ if (send_size[i]){ if (partial_send[i]){ k = start_pos[i] + count[i] - 1; tmp = others_req[i].lens[k]; others_req[i].lens[k] = partial_send[i]; } ompi_datatype_create_hindexed(count[i], &(others_req[i].lens[start_pos[i]]), &(others_req[i].mem_ptrs[start_pos[i]]), MPI_BYTE, &send_type); ompi_datatype_commit(&send_type); ret = MCA_PML_CALL(isend(MPI_BOTTOM, 1, send_type, i, fh->f_rank+i+100*iter, MCA_PML_BASE_SEND_STANDARD, fh->f_comm, requests+nprocs_recv+j)); ompi_datatype_destroy(&send_type); if (partial_send[i]) others_req[i].lens[k] = tmp; j++; } } if (nprocs_recv) { ret = ompi_request_wait_all(nprocs_recv, requests, MPI_STATUS_IGNORE); if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) { two_phase_fill_user_buffer(fh, buf, flat_buf, recv_buf, offset_len, (unsigned *)recv_size, requests, recd_from_proc, contig_access_count, min_st_offset, fd_size, fd_start, fd_end, buftype_extent, striping_unit, aggregator_list); } } ret = ompi_request_wait_all(nprocs_send, requests+nprocs_recv, MPI_STATUS_IGNORE); if (NULL != requests){ free(requests); requests = NULL; } if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)){ for (i=0; i< fh->f_size; i++){ if (recv_size[i]){ free(recv_buf[i]); } } free(recv_buf); } #if TIME_BREAKDOWN end_rcomm_time = MPI_Wtime(); rcomm_time += (end_rcomm_time - start_rcomm_time); #endif exit: return ret; }
int mca_fcoll_static_file_write_all (mca_io_ompio_file_t *fh, void *buf, int count, struct ompi_datatype_t *datatype, ompi_status_public_t *status) { size_t max_data = 0, bytes_per_cycle=0; struct iovec *iov=NULL, *decoded_iov=NULL; uint32_t iov_count=0, iov_index=0; int i=0,j=0,l=0, temp_index; int ret=OMPI_SUCCESS, cycles, local_cycles, *bytes_per_process=NULL; int index, *disp_index=NULL, **blocklen_per_process=NULL; int *iovec_count_per_process=NULL, *displs=NULL; size_t total_bytes_written=0; MPI_Aint **displs_per_process=NULL, *memory_displacements=NULL; MPI_Aint bytes_to_write_in_cycle=0, global_iov_count=0, global_count=0; local_io_array *local_iov_array =NULL, *global_iov_array=NULL; local_io_array *file_offsets_for_agg=NULL; int *sorted=NULL, *sorted_file_offsets=NULL, temp_pindex, *temp_disp_index=NULL; char *send_buf=NULL, *global_buf=NULL; int iov_size=0, current_position=0, *current_index=NULL; int *bytes_remaining=NULL, entries_per_aggregator=0; ompi_datatype_t **recvtype = NULL; MPI_Request *send_req=NULL, *recv_req=NULL; /* For creating datatype of type io_array */ 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; /*----------------------------------------------*/ #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN double write_time = 0.0, start_write_time = 0.0, end_write_time = 0.0; double comm_time = 0.0, start_comm_time = 0.0, end_comm_time = 0.0; double exch_write = 0.0, start_exch = 0.0, end_exch = 0.0; mca_io_ompio_print_entry nentry; #endif #if DEBUG_ON MPI_Aint gc_in; #endif // if (opal_datatype_is_contiguous_memory_layout(&datatype->super,1)) { // fh->f_flags |= OMPIO_CONTIGUOUS_MEMORY; // } /* In case the data is not contigous in memory, decode it into an iovec */ if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) { 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); /* 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; 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; } ompi_datatype_create_struct (3, blocklen, d, types, &io_array_type); ompi_datatype_commit (&io_array_type); /* #########################################################*/ ret = fh->f_generate_current_file_view((struct mca_io_ompio_file_t *)fh, max_data, &iov, &iov_size); if (ret != OMPI_SUCCESS){ fprintf(stderr,"Current File View Generation Error\n"); goto exit; } if (0 == iov_size){ iov_size = 1; } local_iov_array = (local_io_array *)malloc (iov_size * sizeof(local_io_array)); if ( NULL == local_iov_array){ fprintf(stderr,"local_iov_array allocation error\n"); 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; } fh->f_get_bytes_per_agg ( (int *) &bytes_per_cycle); local_cycles = ceil((double)max_data/bytes_per_cycle); 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){ fprintf(stderr,"local cycles allreduce!\n"); goto exit; } if (fh->f_procs_in_group[fh->f_aggregator_index] == 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 *) malloc (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 *) malloc (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 **)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++){ current_index[i] = 0; bytes_remaining[i] =0; blocklen_per_process[i] = NULL; displs_per_process[i] = NULL; } } iovec_count_per_process = (int *) malloc (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 *) 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; } ret = fh->f_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){ fprintf(stderr,"iov size allgatherv array!\n"); goto exit; } if (fh->f_procs_in_group[fh->f_aggregator_index] == 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 (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { global_iov_array = (local_io_array *) malloc (global_iov_count * sizeof(local_io_array)); if (NULL == global_iov_array){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } } ret = fh->f_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 (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { if ( 0 == global_iov_count){ global_iov_count = 1; } 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; } local_heap_sort (global_iov_array, global_iov_count, sorted); } #if DEBUG_ON if (fh->f_procs_in_group[fh->f_aggregator_index] == 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[gc_in].process_id, gc_in, global_iov_array[gc_in].offset, gc_in, global_iov_array[gc_in].length); } } #endif #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_exch = MPI_Wtime(); #endif for (index = 0; index < cycles; index++){ if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { if (NULL == recvtype){ recvtype = (ompi_datatype_t **) malloc (fh->f_procs_per_group * sizeof(ompi_datatype_t *)); if (NULL == recvtype) { 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++){ 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; } } if (local_cycles > index) { if ((index == local_cycles-1) && (max_data % bytes_per_cycle)) { bytes_to_write_in_cycle = max_data % bytes_per_cycle; } else if (max_data <= bytes_per_cycle) { bytes_to_write_in_cycle = max_data; } else { bytes_to_write_in_cycle = bytes_per_cycle; } } else { bytes_to_write_in_cycle = 0; } #if DEBUG_ON /* if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {*/ printf ("***%d: CYCLE %d Bytes %ld**********\n", fh->f_rank, index, bytes_to_write_in_cycle); /* }*/ #endif /********************************************************** **Gather the Data from all the processes at the writers ** *********************************************************/ /* gather from each process how many bytes each will be sending */ fh->f_gather_array (&bytes_to_write_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); /* For each aggregator it needs to get bytes_to_write_in_cycle from each process in group which adds up to bytes_per_cycle */ if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { for (i=0;i<fh->f_procs_per_group; i++){ /* printf("bytes_per_process[%d]: %d\n", i, bytes_per_process[i]); */ #if DEBUG_ON printf ("%d : bytes_per_process : %d\n", fh->f_procs_in_group[i], bytes_per_process[i]); #endif while (bytes_per_process[i] > 0){ if (get_process_id(global_iov_array[sorted[current_index[i]]].process_id, fh) == i){ /* current id owns this entry!*/ /*Add and subtract length and create blocklength and displs array*/ 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; bytes_remaining[i] = 0; disp_index[i] += 1; /* 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] = find_next_index(i, current_index[i], fh, global_iov_array, global_iov_count, sorted); if (current_index[i] == -1){ /* No more entries left, so Its all done! exit!*/ 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] = 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] = 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 = (local_io_array *) malloc(entries_per_aggregator*sizeof(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; 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]; 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; } 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; } 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; } global_count = 0; for (i=0;i<entries_per_aggregator;i++){ temp_pindex = file_offsets_for_agg[sorted_file_offsets[i]].process_id; displs_per_process[temp_pindex][temp_disp_index[temp_pindex]] = memory_displacements[sorted_file_offsets[i]]; if (temp_disp_index[temp_pindex] < disp_index[temp_pindex]) temp_disp_index[temp_pindex] += 1; else{ printf("temp_disp_index[%d]: %d is greater than disp_index[%d]: %d\n", temp_pindex, temp_disp_index[temp_pindex], temp_pindex, disp_index[temp_pindex]); } global_count += file_offsets_for_agg[sorted_file_offsets[i]].length; } if (NULL != temp_disp_index){ free(temp_disp_index); temp_disp_index = NULL; } #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 : %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[ file_offsets_for_agg[sorted_file_offsets[i]].process_id]); } #endif #if DEBUG_ON printf("%d: global_count : %ld, bytes_to_write_in_cycle : %ld, procs_per_group: %d\n", fh->f_rank, global_count, bytes_to_write_in_cycle, fh->f_procs_per_group); #endif #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_comm_time = MPI_Wtime(); #endif global_buf = (char *) malloc (global_count); if (NULL == global_buf){ opal_output(1, "OUT OF MEMORY"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } recv_req = (MPI_Request *) malloc (fh->f_procs_per_group * sizeof(MPI_Request)); if (NULL == recv_req){ 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++){ ompi_datatype_create_hindexed(disp_index[i], blocklen_per_process[i], displs_per_process[i], MPI_BYTE, &recvtype[i]); ompi_datatype_commit(&recvtype[i]); ret = MCA_PML_CALL(irecv(global_buf, 1, recvtype[i], fh->f_procs_in_group[i], 123, fh->f_comm, &recv_req[i])); if (OMPI_SUCCESS != ret){ fprintf(stderr,"irecv Error!\n"); goto exit; } } } if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) { send_buf = &((char*)buf)[total_bytes_written]; } else if (bytes_to_write_in_cycle) { /* allocate a send buffer and copy the data that needs to be sent into it in case the data is non-contigous in memory */ OPAL_PTRDIFF_TYPE mem_address; size_t remaining = 0; size_t temp_position = 0; send_buf = malloc (bytes_to_write_in_cycle); if (NULL == send_buf) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } remaining = bytes_to_write_in_cycle; while (remaining) { mem_address = (OPAL_PTRDIFF_TYPE) (decoded_iov[iov_index].iov_base) + current_position; if (remaining >= (decoded_iov[iov_index].iov_len - current_position)) { memcpy (send_buf+temp_position, (IOVBASE_TYPE *)mem_address, 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 (send_buf+temp_position, (IOVBASE_TYPE *)mem_address, remaining); current_position = current_position + remaining; remaining = 0; } } } total_bytes_written += bytes_to_write_in_cycle; send_req = (MPI_Request *) malloc (sizeof(MPI_Request)); if (NULL == send_req){ opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } ret = MCA_PML_CALL(isend(send_buf, bytes_to_write_in_cycle, MPI_BYTE, fh->f_procs_in_group[fh->f_aggregator_index], 123, MCA_PML_BASE_SEND_STANDARD, fh->f_comm, send_req)); if ( OMPI_SUCCESS != ret ){ fprintf(stderr,"isend error!\n"); goto exit; } ret = ompi_request_wait (send_req, MPI_STATUS_IGNORE); if (OMPI_SUCCESS != ret){ goto exit; } if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { ret = ompi_request_wait_all (fh->f_procs_per_group, recv_req, MPI_STATUS_IGNORE); if (OMPI_SUCCESS != ret){ goto exit; } #if DEBUG_ON printf("************Cycle: %d, Aggregator: %d ***************\n", index+1,fh->f_rank); if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank){ for (i=0 ; i<global_count/4 ; i++) printf (" RECV %d \n",((int *)global_buf)[i]); } #endif } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_comm_time = MPI_Wtime(); comm_time += end_comm_time - start_comm_time; #endif if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { 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; /*First entry for every aggregator*/ fh->f_io_array[fh->f_num_of_io_entries].offset = (IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[0]].offset; fh->f_io_array[fh->f_num_of_io_entries].length = file_offsets_for_agg[sorted_file_offsets[0]].length; fh->f_io_array[fh->f_num_of_io_entries].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_write_time = MPI_Wtime(); #endif if (fh->f_num_of_io_entries) { if ( 0 > fh->f_fbtl->fbtl_pwritev (fh)) { opal_output (1, "WRITE FAILED\n"); ret = OMPI_ERROR; goto exit; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_write_time = MPI_Wtime(); write_time += end_write_time - start_write_time; #endif } if (NULL != send_req){ free(send_req); send_req = NULL; } if (fh->f_procs_in_group[fh->f_aggregator_index] == 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; } for (i = 0; i < fh->f_procs_per_group; i++) ompi_datatype_destroy(recvtype+i); if (NULL != recvtype){ free(recvtype); recvtype=NULL; } if (NULL != recv_req){ free(recv_req); recv_req = NULL; } if (NULL != global_buf) { free (global_buf); global_buf = NULL; } } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_exch = MPI_Wtime(); exch_write += end_exch - start_exch; nentry.time[0] = write_time; nentry.time[1] = comm_time; nentry.time[2] = exch_write; if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) nentry.aggregator = 1; else nentry.aggregator = 0; nentry.nprocs_for_coll = static_num_io_procs; if (!fh->f_full_print_queue(WRITE_PRINT_QUEUE)){ fh->f_register_print_entry(WRITE_PRINT_QUEUE, nentry); } #endif exit: if (NULL != decoded_iov){ free(decoded_iov); decoded_iov = NULL; } if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) { if (NULL != local_iov_array){ free(local_iov_array); local_iov_array = NULL; } 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; } if (NULL != displs_per_process[l]){ free(displs_per_process[l]); displs_per_process[l] = NULL; } } } if (NULL != send_buf){ free(send_buf); send_buf = NULL; } if (NULL != global_buf){ free(global_buf); global_buf = NULL; } if (NULL != recvtype){ free(recvtype); recvtype = 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 != displs_per_process){ free(displs_per_process); displs_per_process = NULL; } if (NULL != blocklen_per_process){ free(blocklen_per_process); blocklen_per_process = NULL; } if(NULL != current_index){ free(current_index); current_index = NULL; } if(NULL != bytes_remaining){ free(bytes_remaining); bytes_remaining = NULL; } if (NULL != disp_index){ free(disp_index); disp_index = NULL; } if (NULL != sorted) { free(sorted); sorted = NULL; } return ret; }
int mca_fcoll_dynamic_file_write_all (ompio_file_t *fh, const void *buf, int count, struct ompi_datatype_t *datatype, ompi_status_public_t *status) { MPI_Aint total_bytes_written = 0; /* total bytes that have been written*/ MPI_Aint total_bytes = 0; /* total bytes to be written */ MPI_Aint bytes_to_write_in_cycle = 0; /* left to be written in a cycle*/ MPI_Aint bytes_per_cycle = 0; /* total written in each cycle by each process*/ int index = 0; 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 written from the current value from total_bytes_per_process */ int bytes_sent = 0, ret =0; int blocks=0, entries_per_aggregator=0; /* iovec structure and count of the buffer passed in */ uint32_t iov_count = 0; struct iovec *decoded_iov = NULL; int iov_index = 0; char *send_buf = NULL; size_t current_position = 0; struct iovec *local_iov_array=NULL, *global_iov_array=NULL; mca_io_ompio_local_io_array *file_offsets_for_agg=NULL; /* global iovec at the writers that contain the iovecs created from file_set_view */ uint32_t total_fview_count = 0; int local_count = 0, temp_pindex; int *fview_count = NULL, *disp_index=NULL, *temp_disp_index=NULL; int current_index = 0, temp_index=0; char *global_buf = NULL; MPI_Aint global_count = 0; /* array that contains the sorted indices of the global_iov */ int *sorted = NULL, *sorted_file_offsets=NULL; int *displs = NULL; int dynamic_num_io_procs; size_t max_data = 0, datatype_size = 0; int **blocklen_per_process=NULL; MPI_Aint **displs_per_process=NULL, *memory_displacements=NULL; ompi_datatype_t **recvtype = NULL; MPI_Aint *total_bytes_per_process = NULL; MPI_Request send_req=NULL, *recv_req=NULL; int my_aggregator=-1; bool sendbuf_is_contiguous = false; size_t ftype_size; ptrdiff_t ftype_extent, lb; #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN double write_time = 0.0, start_write_time = 0.0, end_write_time = 0.0; double comm_time = 0.0, start_comm_time = 0.0, end_comm_time = 0.0; double exch_write = 0.0, start_exch = 0.0, end_exch = 0.0; mca_common_ompio_print_entry nentry; #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 == (ptrdiff_t) ftype_size) && opal_datatype_is_contiguous_memory_layout(&datatype->super,1) && 0 == lb ) { sendbuf_is_contiguous = true; } if (! sendbuf_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; } dynamic_num_io_procs = fh->f_get_mca_parameter_value ( "num_aggregators", strlen ("num_aggregators")); if ( OMPI_ERR_MAX == dynamic_num_io_procs ) { ret = OMPI_ERROR; goto exit; } ret = mca_common_ompio_set_aggregator_props ((struct ompio_file_t *) fh, dynamic_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_comm_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_comm_time = MPI_Wtime(); comm_time += (end_comm_time - start_comm_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 local offsets/lengths array corresponding to *** this write operation ********************************************************************/ 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; } #if DEBUG_ON for (i=0 ; i<local_count ; i++) { printf("%d: OFFSET: %d LENGTH: %ld\n", fh->f_rank, local_iov_array[i].iov_base, local_iov_array[i].iov_len); } #endif /************************************************************* *** 4. Allgather the offset/lengths array from 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_comm_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_comm_time = MPI_Wtime(); comm_time += (end_comm_time - start_comm_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 printf("total_fview_count : %d\n", total_fview_count); 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_comm_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_comm_time = MPI_Wtime(); comm_time += (end_comm_time - start_comm_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 (NULL != displs){ free(displs); displs=NULL; } #if DEBUG_ON if (my_aggregator == fh->f_rank) { uint32_t tv=0; for (tv=0 ; tv<total_fview_count ; tv++) { printf("%d: OFFSET: %lld LENGTH: %ld\n", fh->f_rank, global_iov_array[sorted[tv]].iov_base, global_iov_array[sorted[tv]].iov_len); } } #endif /************************************************************* *** 6. Determine the number of cycles required to execute this *** operation *************************************************************/ bytes_per_cycle = fh->f_bytes_per_agg; 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 **)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; } recv_req = (MPI_Request *)malloc ((fh->f_procs_per_group)*sizeof(MPI_Request)); if ( NULL == recv_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"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } recvtype = (ompi_datatype_t **) malloc (fh->f_procs_per_group * sizeof(ompi_datatype_t *)); if (NULL == recvtype) { 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++){ recvtype[l] = MPI_DATATYPE_NULL; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_exch = 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 != recvtype){ for (i =0; i< fh->f_procs_per_group; i++) { if ( MPI_DATATYPE_NULL != recvtype[i] ) { ompi_datatype_destroy(&recvtype[i]); recvtype[i] = MPI_DATATYPE_NULL; } } } for(l=0;l<fh->f_procs_per_group;l++){ disp_index[l] = 1; free(blocklen_per_process[l]); free(displs_per_process[l]); blocklen_per_process[l] = (int *) calloc (1, sizeof(int)); displs_per_process[l] = (MPI_Aint *) calloc (1, sizeof(MPI_Aint)); if (NULL == displs_per_process[l] || NULL == blocklen_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 written in this cycle **************************************************************************/ if (cycles-1 == index) { bytes_to_write_in_cycle = total_bytes - bytes_per_cycle*index; } else { bytes_to_write_in_cycle = bytes_per_cycle; } #if DEBUG_ON if (my_aggregator == fh->f_rank) { printf ("****%d: CYCLE %d Bytes %lld**********\n", fh->f_rank, index, bytes_to_write_in_cycle); } #endif /********************************************************** **Gather the Data from all the processes at the writers ** *********************************************************/ #if DEBUG_ON printf("bytes_to_write_in_cycle: %ld, cycle : %d\n", bytes_to_write_in_cycle, index); #endif /***************************************************************** *** 7c. Calculate how much data will be contributed in this cycle *** by each process *****************************************************************/ bytes_sent = 0; /* The blocklen and displs calculation only done at aggregators!*/ while (bytes_to_write_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_write_in_cycle) { /* The 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); /* In this cases the length is consumed so allocating for next displacement and blocklength*/ 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_sent += bytes_remaining; } current_index ++; bytes_to_write_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_write_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_sent += bytes_to_write_in_cycle; } bytes_remaining -= bytes_to_write_in_cycle; bytes_to_write_in_cycle = 0; break; } } else { /* No partially used entry available, have to start a new one */ if (bytes_to_write_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_write_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_sent += bytes_to_write_in_cycle; } bytes_remaining = global_iov_array[sorted[current_index]].iov_len - bytes_to_write_in_cycle; bytes_to_write_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; /*realloc for next blocklength and assign this displacement and check for next displs as the total length of this entry has been consumed!*/ 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_sent += global_iov_array[sorted[current_index]].iov_len; } bytes_to_write_in_cycle -= global_iov_array[sorted[current_index]].iov_len; current_index ++; continue; } } } /************************************************************************* *** 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 DEBUG_ON printf("%d: cycle: %d, bytes_sent: %d\n ",fh->f_rank,index, bytes_sent); printf("%d : Entries per aggregator : %d\n",fh->f_rank,entries_per_aggregator); #endif 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; 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]; file_offsets_for_agg[temp_index].process_id = i; file_offsets_for_agg[temp_index].offset = displs_per_process[i][j]; temp_index++; #if DEBUG_ON printf("************Cycle: %d, Aggregator: %d ***************\n", index+1,fh->f_rank); 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 } } } } else{ continue; } /* Sort the displacements for each aggregator*/ local_heap_sort (file_offsets_for_agg, entries_per_aggregator, sorted_file_offsets); /*create contiguous memory displacements based on blocklens on the same displs array and map it to this aggregator's actual file-displacements (this is in the io-array created above)*/ 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; } 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; } /*Now update the displacements array with memory offsets*/ global_count = 0; for (i=0;i<entries_per_aggregator;i++){ temp_pindex = file_offsets_for_agg[sorted_file_offsets[i]].process_id; displs_per_process[temp_pindex][temp_disp_index[temp_pindex]] = memory_displacements[sorted_file_offsets[i]]; if (temp_disp_index[temp_pindex] < disp_index[temp_pindex]) temp_disp_index[temp_pindex] += 1; else{ printf("temp_disp_index[%d]: %d is greater than disp_index[%d]: %d\n", temp_pindex, temp_disp_index[temp_pindex], temp_pindex, disp_index[temp_pindex]); } global_count += file_offsets_for_agg[sorted_file_offsets[i]].length; } if (NULL != temp_disp_index){ free(temp_disp_index); temp_disp_index = NULL; } #if DEBUG_ON printf("************Cycle: %d, Aggregator: %d ***************\n", index+1,fh->f_rank); 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){ 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); } } } 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\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]]); } printf("%d : global_count : %ld, bytes_sent : %d\n", fh->f_rank,global_count, bytes_sent); #endif #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_comm_time = MPI_Wtime(); #endif /************************************************************************* *** 7e. Perform the actual communication *************************************************************************/ for (i=0;i<fh->f_procs_per_group; i++) { recv_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, &recvtype[i]); ompi_datatype_commit(&recvtype[i]); opal_datatype_type_size(&recvtype[i]->super, &datatype_size); if (datatype_size){ ret = MCA_PML_CALL(irecv(global_buf, 1, recvtype[i], fh->f_procs_in_group[i], 123, fh->f_comm, &recv_req[i])); if (OMPI_SUCCESS != ret){ goto exit; } } } } } /* end if (my_aggregator == fh->f_rank ) */ if ( sendbuf_is_contiguous ) { send_buf = &((char*)buf)[total_bytes_written]; } else if (bytes_sent) { /* allocate a send buffer and copy the data that needs to be sent into it in case the data is non-contigous in memory */ ptrdiff_t mem_address; size_t remaining = 0; size_t temp_position = 0; send_buf = malloc (bytes_sent); if (NULL == send_buf) { opal_output (1, "OUT OF MEMORY\n"); ret = OMPI_ERR_OUT_OF_RESOURCE; goto exit; } remaining = bytes_sent; 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 (send_buf+temp_position, (IOVBASE_TYPE *)mem_address, 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 (send_buf+temp_position, (IOVBASE_TYPE *) mem_address, remaining); current_position = current_position + remaining; remaining = 0; } } } total_bytes_written += bytes_sent; /* Gather the sendbuf from each process in appropritate locations in aggregators*/ if (bytes_sent){ ret = MCA_PML_CALL(isend(send_buf, bytes_sent, MPI_BYTE, my_aggregator, 123, MCA_PML_BASE_SEND_STANDARD, fh->f_comm, &send_req)); if ( OMPI_SUCCESS != ret ){ goto exit; } ret = ompi_request_wait(&send_req, MPI_STATUS_IGNORE); if (OMPI_SUCCESS != ret){ goto exit; } } if (my_aggregator == fh->f_rank) { ret = ompi_request_wait_all (fh->f_procs_per_group, recv_req, MPI_STATUS_IGNORE); if (OMPI_SUCCESS != ret){ goto exit; } } #if DEBUG_ON if (my_aggregator == fh->f_rank){ printf("************Cycle: %d, Aggregator: %d ***************\n", index+1,fh->f_rank); for (i=0 ; i<global_count/4 ; i++) printf (" RECV %d \n",((int *)global_buf)[i]); } #endif if (! sendbuf_is_contiguous) { if (NULL != send_buf) { free (send_buf); send_buf = NULL; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_comm_time = MPI_Wtime(); comm_time += (end_comm_time - start_comm_time); #endif /********************************************************** *** 7f. Create the io array, and pass it to fbtl *********************************************************/ if (my_aggregator == fh->f_rank) { #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN start_write_time = MPI_Wtime(); #endif 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; /*First entry for every aggregator*/ 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 the enrties are contiguous merge them, else make a new entry */ 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, (ptrdiff_t)fh->f_io_array[i].offset, fh->f_io_array[i].length); } #endif if (fh->f_num_of_io_entries) { if ( 0 > fh->f_fbtl->fbtl_pwritev (fh)) { opal_output (1, "WRITE FAILED\n"); ret = OMPI_ERROR; goto exit; } } #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_write_time = MPI_Wtime(); write_time += end_write_time - start_write_time; #endif } /* end if (my_aggregator == fh->f_rank) */ } /* end for (index = 0; index < cycles; index++) */ #if OMPIO_FCOLL_WANT_TIME_BREAKDOWN end_exch = MPI_Wtime(); exch_write += end_exch - start_exch; nentry.time[0] = write_time; nentry.time[1] = comm_time; nentry.time[2] = exch_write; if (my_aggregator == fh->f_rank) nentry.aggregator = 1; else nentry.aggregator = 0; nentry.nprocs_for_coll = dynamic_num_io_procs; if (!mca_common_ompio_full_print_queue(fh->f_coll_write_time)){ mca_common_ompio_register_print_entry(fh->f_coll_write_time, nentry); } #endif exit : 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 != recvtype){ for (i =0; i< fh->f_procs_per_group; i++) { if ( MPI_DATATYPE_NULL != recvtype[i] ) { ompi_datatype_destroy(&recvtype[i]); } } free(recvtype); recvtype=NULL; } if (NULL != fh->f_io_array) { free (fh->f_io_array); fh->f_io_array = NULL; } if (NULL != disp_index){ free(disp_index); disp_index = NULL; } if (NULL != recvtype){ free(recvtype); recvtype=NULL; } if (NULL != recv_req){ free(recv_req); recv_req = NULL; } if (NULL != global_buf) { free (global_buf); global_buf = NULL; } for(l=0;l<fh->f_procs_per_group;l++){ if (NULL != blocklen_per_process){ free(blocklen_per_process[l]); } if (NULL != displs_per_process){ free(displs_per_process[l]); } } free(blocklen_per_process); free(displs_per_process); } if (NULL != displs){ free(displs); displs=NULL; } if (! sendbuf_is_contiguous) { if (NULL != send_buf) { free (send_buf); send_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; } return OMPI_SUCCESS; }
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; }