diff options
Diffstat (limited to 'src/H5Dmpio.c')
| -rw-r--r-- | src/H5Dmpio.c | 276 |
1 files changed, 138 insertions, 138 deletions
diff --git a/src/H5Dmpio.c b/src/H5Dmpio.c index 7ad8e50..ead1777 100644 --- a/src/H5Dmpio.c +++ b/src/H5Dmpio.c @@ -104,29 +104,29 @@ typedef struct H5D_common_coll_info_t { size_t mpi_buf_count; haddr_t chunk_addr; } H5D_common_coll_info_t; - + /********************/ /* Local Prototypes */ /********************/ -static herr_t -H5D_multi_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, +static herr_t +H5D_multi_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, hbool_t do_write); static herr_t -H5D_link_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, +H5D_link_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, hbool_t do_write,int sum_chunk); -static herr_t +static herr_t H5D_inter_collective_io(H5D_io_info_t *io_info,const H5S_t *file_space, - const H5S_t *mem_space,haddr_t addr, + const H5S_t *mem_space,haddr_t addr, const void *buf, hbool_t do_write ); -static herr_t +static herr_t H5D_final_collective_io(H5D_io_info_t *io_info,MPI_Datatype*mpi_file_type, MPI_Datatype *mpi_buf_type, - H5D_common_coll_info_t* coll_info, + H5D_common_coll_info_t* coll_info, const void *buf, hbool_t do_write); #ifdef OLD_WAY static herr_t @@ -134,14 +134,14 @@ H5D_pre_sort_chunk(H5D_io_info_t *io_info,int total_chunks, haddr_t total_chunk_addr_array[]); #endif -static herr_t +static herr_t H5D_sort_chunk(H5D_io_info_t * io_info, fm_map *fm, H5D_chunk_addr_info_t chunk_addr_info_array[], int many_chunk_opt); -static herr_t -H5D_obtain_mpio_mode(H5D_io_info_t* io_info, +static herr_t +H5D_obtain_mpio_mode(H5D_io_info_t* io_info, fm_map *fm, uint8_t assign_io_mode[], haddr_t chunk_addr[]); @@ -151,7 +151,7 @@ static herr_t H5D_ioinfo_make_coll(H5D_io_info_t *io_info); static herr_t H5D_mpio_get_min_chunk(const H5D_io_info_t *io_info, const fm_map *fm, int *min_chunkf); static int H5D_cmp_chunk_addr(const void *addr1, const void *addr2); -static herr_t +static herr_t H5D_mpio_get_sum_chunk(const H5D_io_info_t *io_info, const fm_map *fm, int *sum_chunkf); @@ -284,10 +284,10 @@ done: * Decription: If H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS is not defined, collective IO with no contribution from one or more processes are not assured. We will check the minimum - number of chunks the process is used. If the number is + number of chunks the process is used. If the number is zero, we will use independent IO mode instead. This is necessary with Linked chunk IO. - * Purpose: Checks if it is possible to do collective IO + * Purpose: Checks if it is possible to do collective IO * * Return: Success: Non-negative: TRUE or FALSE * Failure: Negative @@ -307,8 +307,8 @@ H5D_mpio_chunk_adjust_iomode(H5D_io_info_t *io_info, const fm_map *fm) { #ifndef H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS if(H5D_mpio_get_min_chunk(io_info,fm,&min_chunk)<0) - HGOTO_ERROR(H5E_DATASPACE, H5E_CANTSWAP, FAIL, "unable to obtain the min chunk number of all processes"); - if(min_chunk == 0) { + HGOTO_ERROR(H5E_DATASPACE, H5E_CANTSWAP, FAIL, "unable to obtain the min chunk number of all processes"); + if(min_chunk == 0) { H5P_genplist_t *dx_plist; /* Data transer property list */ /* Get the dataset transfer property list */ @@ -347,11 +347,11 @@ done: */ herr_t H5D_mpio_select_read(H5D_io_info_t *io_info, - size_t mpi_buf_count, + size_t mpi_buf_count, const size_t UNUSED elmt_size, - const H5S_t UNUSED *file_space, + const H5S_t UNUSED *file_space, const H5S_t UNUSED *mem_space, - haddr_t addr, + haddr_t addr, void *buf/*out*/) { herr_t ret_value = SUCCEED; @@ -378,9 +378,9 @@ done: */ herr_t H5D_mpio_select_write(H5D_io_info_t *io_info, - size_t mpi_buf_count, + size_t mpi_buf_count, const size_t UNUSED elmt_size, - const H5S_t UNUSED *file_space, + const H5S_t UNUSED *file_space, const H5S_t UNUSED *mem_space, haddr_t addr, const void *buf) @@ -569,8 +569,8 @@ done: * Function: H5D_contig_collective_io * * Purpose: Wrapper Routine for H5D_inter_collective_io - The starting address of contiguous storage is passed - * + The starting address of contiguous storage is passed + * * * Return: Non-negative on success/Negative on failure * @@ -581,11 +581,11 @@ done: *------------------------------------------------------------------------- */ herr_t -H5D_contig_collective_io(H5D_io_info_t *io_info, +H5D_contig_collective_io(H5D_io_info_t *io_info, const H5S_t *file_space, const H5S_t *mem_space, const void *buf, - hbool_t do_write) + hbool_t do_write) { @@ -607,8 +607,8 @@ H5D_contig_collective_io(H5D_io_info_t *io_info, #endif if(H5D_inter_collective_io(io_info,file_space,mem_space,addr,buf,do_write)<0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL,"couldn't finish shared collective MPI-IO"); - - done: + + done: FUNC_LEAVE_NOAPI(ret_value) } /* end H5D_contig_collective_io */ @@ -616,10 +616,10 @@ H5D_contig_collective_io(H5D_io_info_t *io_info, /*------------------------------------------------------------------------- * Function: H5D_chunk_collective_io * - * Purpose: Routine for - 1) choose an IO option: + * Purpose: Routine for + 1) choose an IO option: a) One collective IO defined by one MPI derived datatype to link through all chunks - or b) multiple chunk IOs,to do MPI-IO for each chunk, the IO mode may be adjusted + or b) multiple chunk IOs,to do MPI-IO for each chunk, the IO mode may be adjusted due to the selection pattern for each chunk. * For option a) 1. Sort the chunk address, obtain chunk info according to the sorted chunk address @@ -633,7 +633,7 @@ H5D_contig_collective_io(H5D_io_info_t *io_info, 2. Depending on whether the IO mode is collective or independent or none, Create either MPI derived datatype for each chunk to do collective IO or just do independent IO 3. Set up collective IO property list for collective mode - 4. DO IO + 4. DO IO * * Return: Non-negative on success/Negative on failure * @@ -643,44 +643,44 @@ H5D_contig_collective_io(H5D_io_info_t *io_info, * *------------------------------------------------------------------------- */ -herr_t -H5D_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, hbool_t do_write) +herr_t +H5D_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, hbool_t do_write) { int io_option = H5D_MULTI_CHUNK_IO_MORE_OPT; int sum_chunk = 0,mpi_size; unsigned one_link_chunk_io_threshold; - H5P_genplist_t *plist; + H5P_genplist_t *plist; H5FD_mpio_chunk_opt_t chunk_opt_mode; #ifdef H5_HAVE_INSTRUMENTED_LIBRARY htri_t check_prop,temp_not_link_io = FALSE; int prop_value,new_value; #endif - herr_t ret_value = SUCCEED; + herr_t ret_value = SUCCEED; FUNC_ENTER_NOAPI_NOINIT(H5D_chunk_collective_io) assert (IS_H5FD_MPIO(io_info->dset->oloc.file)); - + /* Obtain the data transfer properties */ if(NULL == (plist = H5I_object(io_info->dxpl_id))) HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a file access property list") - + /* Check the optional property list on what to do with collective chunk IO. */ chunk_opt_mode=(H5FD_mpio_chunk_opt_t)H5P_peek_unsigned(plist,H5D_XFER_MPIO_CHUNK_OPT_HARD_NAME); #ifdef KENT printf("chunk_opt_mode = %d\n",chunk_opt_mode); #endif - + if(chunk_opt_mode == H5FD_MPIO_CHUNK_ONE_IO) io_option = H5D_ONE_LINK_CHUNK_IO;/*no opt*/ else if(chunk_opt_mode == H5FD_MPIO_CHUNK_MULTI_IO) io_option = H5D_MULTI_CHUNK_IO;/*no opt */ else { - if(H5D_mpio_get_sum_chunk(io_info,fm,&sum_chunk)<0) - HGOTO_ERROR(H5E_DATASPACE, H5E_CANTSWAP, FAIL, "unable to obtain the total chunk number of all processes"); + if(H5D_mpio_get_sum_chunk(io_info,fm,&sum_chunk)<0) + HGOTO_ERROR(H5E_DATASPACE, H5E_CANTSWAP, FAIL, "unable to obtain the total chunk number of all processes"); if((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file))<0) HGOTO_ERROR (H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size"); - + if(NULL == (plist = H5I_object(io_info->dxpl_id))) HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a file access property list") @@ -732,8 +732,8 @@ H5D_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, hbool HGOTO_ERROR(H5E_PLIST, H5E_UNSUPPORTED, FAIL, "unable to get property value"); } } - - + + #endif #ifndef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS if(io_option == H5D_ONE_LINK_CHUNK_IO ) io_option = H5D_MULTI_CHUNK_IO ;/* We can not do this with one chunk IO. */ @@ -745,9 +745,9 @@ H5D_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, hbool if(H5D_link_chunk_collective_io(io_info,fm,buf,do_write,sum_chunk)<0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL,"couldn't finish linked chunk MPI-IO"); } - + else { /*multiple chunk IOs without opt */ - + if(H5D_multi_chunk_collective_io(io_info,fm,buf,do_write)<0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL,"couldn't finish multiple chunk MPI-IO"); @@ -764,7 +764,7 @@ done: 1. Sort the chunk address and chunk info 2. Build up MPI derived datatype for each chunk 3. Build up the final MPI derived datatype - 4. Use common collective IO routine to do MPI-IO + 4. Use common collective IO routine to do MPI-IO * * Return: Non-negative on success/Negative on failure @@ -788,9 +788,9 @@ H5D_link_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, size_t mpi_buf_count; size_t mpi_file_count; hbool_t mbt_is_derived=0, /* Whether the buffer (memory) type is derived and needs to be free'd */ - mft_is_derived=0; /* Whether the file type is derived and needs to be free'd */ - - int mpi_size,mpi_code; /* MPI return code */ + mft_is_derived=0; /* Whether the file type is derived and needs to be free'd */ + + int mpi_size,mpi_code; /* MPI return code */ int i,num_chunk=0,total_chunks; size_t ori_num_chunk; @@ -818,8 +818,8 @@ H5D_link_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, char *bc_percent = NULL; char *bcc_percent = NULL; #endif - herr_t ret_value = SUCCEED; - + herr_t ret_value = SUCCEED; + FUNC_ENTER_NOAPI_NOINIT(H5D_link_chunk_collective_io) ori_total_chunks = fm->total_chunks; H5_ASSIGN_OVERFLOW(total_chunks,ori_total_chunks,hsize_t,int); @@ -829,7 +829,7 @@ H5D_link_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, H5SL_node_t *chunk_node; H5D_chunk_info_t *chunk_info; H5D_storage_t store; - + chunk_node = H5SL_first(fm->fsel); if(chunk_node == NULL) { if(H5D_istore_chunkmap(io_info,total_chunks,&chunk_base_addr,fm->down_chunks)<0) @@ -846,7 +846,7 @@ H5D_link_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, if(HADDR_UNDEF==(chunk_base_addr = H5D_istore_get_addr(io_info,NULL))) HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL,"couldn't get chunk info from skipped list"); - + #ifdef KENT printf("before inter_collective_io for total chunk = 1 \n"); #endif @@ -864,7 +864,7 @@ printf("before inter_collective_io for total chunk = 1 \n"); printf("total_chunks = %d\n",(int)total_chunks); #endif - + if(num_chunk == 0) total_chunk_addr_array = H5MM_malloc(sizeof(haddr_t)*total_chunks); else { @@ -888,12 +888,12 @@ printf("total_chunks = %d\n",(int)total_chunks); "bc" means 'b-tree iterately obtain all chunk addresses individually', the default one means 'obtaining the chunk address individually', */ - + if(bcc_percent=getenv("BCC_PERCENT")){ bsearch_coll_chunk_threshold = atoi(bcc_percent); assert((bsearch_coll_chunk_threshold >=0) &&(bsearch_coll_chunk_threshold <=100)); } - else + else bsearch_coll_chunk_threshold = H5D_ALL_CHUNK_ADDR_THRES_COL; #else bsearch_coll_chunk_threshold = H5D_ALL_CHUNK_ADDR_THRES_COL; /*This number may be changed according to the performance study */ @@ -902,9 +902,9 @@ printf("total_chunks = %d\n",(int)total_chunks); if((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file))<0) HGOTO_ERROR (H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size"); - /* Calculate the actual threshold to obtain all chunk addresses collectively + /* Calculate the actual threshold to obtain all chunk addresses collectively The bigger this number is, the more possible the use of obtaining chunk address collectively. */ - /* For non-optimization one-link IO, + /* For non-optimization one-link IO, actual bsearch threshold is always 0, we would always want to obtain the chunk addresses individually for each process. */ @@ -921,7 +921,7 @@ printf("total_chunks = %d\n",(int)total_chunks); bsearch_chunk_ratio = atoi(bc_percent); assert((bsearch_chunk_ratio<=100)&&(bsearch_chunk_ratio>=0)); } - else + else bsearch_chunk_ratio = H5D_ALL_CHUNK_ADDR_THRES_IND; #else bsearch_chunk_ratio = H5D_ALL_CHUNK_ADDR_THRES_IND; /*This number may be changed according to the performance study */ @@ -931,20 +931,20 @@ printf("total_chunks = %d\n",(int)total_chunks); The unit of the threshold is the number of chunks. The value should be at least 1. It can be calculated as follows: - if(total_chunks*bsearch_chunk_ratio/100 <=1) + if(total_chunks*bsearch_chunk_ratio/100 <=1) bsearch_chunk_threahold = 1; - else + else bsearch_chunk_threshold = total_chunks*bsearch_chunk_ratio/100; - In order to make the caluculation more efficient, + In order to make the caluculation more efficient, we use the following approximate formula to calculate the threshold. bsearch_chunk_threshold = 1+ (total_chunks*bsearch_chunk_ratio-99)/100; The only difference is when total_chunks* besearch_chunk_ratio == 100n+99; - the approximate formula will give value (n+1) instead of n for threshold. + the approximate formula will give value (n+1) instead of n for threshold. That shouldn't matter much from our persective. - */ - + */ + bsearch_chunk_threshold = 1 +(total_chunks*bsearch_chunk_ratio-99)/100; if(num_chunk > bsearch_chunk_threshold) many_chunk_opt = H5D_OBTAIN_ALL_CHUNK_ADDR_IND; if((sum_chunk == 0) && (total_chunks >= H5D_ALL_CHUNK_ADDR_THRES_IND_NUM)) @@ -953,14 +953,14 @@ printf("total_chunks = %d\n",(int)total_chunks); #ifdef KENT printf("before sorting the chunk address \n"); #endif - /* Sort the chunk address + /* Sort the chunk address when chunk optimization selection is either H5D_OBTAIN_*/ if(num_chunk == 0){ /* special case: this process doesn't select anything */ if(H5D_istore_chunkmap(io_info,total_chunks,total_chunk_addr_array,fm->down_chunks)<0) HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get chunk address"); chunk_base_addr = total_chunk_addr_array[0]; } - + else { if(H5D_sort_chunk(io_info,fm,chunk_addr_info_array,many_chunk_opt)<0) HGOTO_ERROR (H5E_DATASPACE, H5E_CANTSWAP, FAIL, "unable to sort chunk address"); @@ -969,8 +969,8 @@ printf("before sorting the chunk address \n"); #ifdef KENT printf("after sorting the chunk address \n"); #endif - - /* Obtain MPI derived datatype from all individual chunks */ + + /* Obtain MPI derived datatype from all individual chunks */ for ( i = 0; i < num_chunk; i++) { /* Disk MPI derived datatype */ if(H5S_mpio_space_type(chunk_addr_info_array[i].chunk_info.fspace,src_type_size,&chunk_ftype[i], @@ -981,7 +981,7 @@ printf("after sorting the chunk address \n"); if(H5S_mpio_space_type(chunk_addr_info_array[i].chunk_info.mspace,dst_type_size,&chunk_mtype[i], &mpi_buf_count,&mpi_buf_extra_offset,&mbt_is_derived)<0) HGOTO_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"couldn't create MPI buf type"); - + /* Chunk address relative to the first chunk */ chunk_addr_info_array[i].chunk_addr -= chunk_base_addr; H5_ASSIGN_OVERFLOW(chunk_disp_array[i],chunk_addr_info_array[i].chunk_addr,haddr_t,MPI_Aint); @@ -989,7 +989,7 @@ printf("after sorting the chunk address \n"); blocklen_value = 1; if(num_chunk){ - + /* initialize the buffer with the constant value 1 */ H5V_array_fill(blocklen,&blocklen_value,sizeof(int),(size_t)num_chunk); @@ -1022,7 +1022,7 @@ printf("after sorting the chunk address \n"); else {/* no selection at all for this process */ chunk_final_ftype = MPI_BYTE; chunk_final_mtype = MPI_BYTE; - + /* buffer, file derived datatypes should be true */ coll_info.mbt_is_derived = 0; coll_info.mft_is_derived = 0; @@ -1032,7 +1032,7 @@ printf("after sorting the chunk address \n"); #ifdef KENT printf("before coming to final collective IO\n"); #endif - + if(H5D_final_collective_io(io_info,&chunk_final_ftype,&chunk_final_mtype,&coll_info,buf,do_write)<0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL,"couldn't finish MPI-IO"); @@ -1042,7 +1042,7 @@ printf("before freeing memory inside H5D_link_collective_io ret_value = %d\n",r #endif - + if (fm->total_chunks != 1) { if(num_chunk == 0) HDfree(total_chunk_addr_array); else { @@ -1072,7 +1072,7 @@ printf("before leaving H5D_link_collective_io ret_value = %d\n",ret_value); 1. Use MPI_gather and MPI_Bcast to obtain IO mode in each chunk(collective/independent/none) 2. Depending on whether the IO mode is collective or independent or none, Create either MPI derived datatype for each chunk or just do independent IO - 3. Use common collective IO routine to do MPI-IO + 3. Use common collective IO routine to do MPI-IO * * Return: Non-negative on success/Negative on failure * @@ -1082,8 +1082,8 @@ printf("before leaving H5D_link_collective_io ret_value = %d\n",ret_value); * *------------------------------------------------------------------------- */ -static herr_t -H5D_multi_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, hbool_t do_write) +static herr_t +H5D_multi_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, hbool_t do_write) { int i,total_chunk; @@ -1096,7 +1096,7 @@ H5D_multi_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, H5D_storage_t store; /* union of EFL and chunk pointer in file space */ hbool_t select_chunk; hbool_t last_io_mode_coll = TRUE; - herr_t ret_value = SUCCEED; + herr_t ret_value = SUCCEED; #ifdef KENT int mpi_rank; #endif @@ -1117,7 +1117,7 @@ H5D_multi_chunk_collective_io(H5D_io_info_t *io_info,fm_map *fm,const void *buf, #endif /* obtain IO option for each chunk */ - if(H5D_obtain_mpio_mode(io_info,fm,chunk_io_option,chunk_addr)<0) + if(H5D_obtain_mpio_mode(io_info,fm,chunk_io_option,chunk_addr)<0) HGOTO_ERROR (H5E_DATASET, H5E_CANTRECV, FAIL, "unable to obtain MPIO mode"); for( i = 0; i<total_chunk;i++){ @@ -1140,7 +1140,7 @@ printf("mpi_rank = %d, chunk index = %d\n",mpi_rank,i); if(NULL ==(chunk_node = H5SL_first(fm->fsel))) HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL,"couldn't get chunk node from skipped list"); #else - + if(NULL ==(chunk_node = H5SL_first(fm->fsel))) HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL,"couldn't get chunk node from skipped list"); while(chunk_node){ @@ -1161,30 +1161,30 @@ printf("mpi_rank = %d, chunk index = %d\n",mpi_rank,i); #endif } - if(chunk_io_option[i] == 1){ /*collective IO for this chunk, + if(chunk_io_option[i] == 1){ /*collective IO for this chunk, note: even there is no selection for this process, the process still needs to contribute MPI NONE TYPE.*/ #ifdef KENT printf("inside collective chunk IO mpi_rank = %d, chunk index = %d\n",mpi_rank,i); #endif - + if(!last_io_mode_coll) /* Switch back to collective I/O */ if(H5D_ioinfo_make_coll(io_info) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't switch to collective I/O") - + if(select_chunk){ if(H5D_inter_collective_io(io_info,chunk_info->fspace,chunk_info->mspace, chunk_addr[i],buf,do_write )<0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL,"couldn't finish shared collective MPI-IO"); - + } else{ if(H5D_inter_collective_io(io_info,NULL,NULL, chunk_addr[i],buf,do_write )<0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL,"couldn't finish shared collective MPI-IO"); - - } + + } last_io_mode_coll = TRUE; } @@ -1192,7 +1192,7 @@ printf("inside collective chunk IO mpi_rank = %d, chunk index = %d\n",mpi_rank,i #ifdef KENT printf("inside independent IO mpi_rank = %d, chunk index = %d\n",mpi_rank,i); #endif - + HDassert(chunk_io_option[i] == 0); if(!select_chunk) continue; /* this process has nothing to do with this chunk, continue! */ if(last_io_mode_coll) @@ -1206,16 +1206,16 @@ printf("inside independent IO mpi_rank = %d, chunk index = %d\n",mpi_rank,i); chunk_info->fspace,chunk_info->mspace,0, buf); /* Check return value of the write */ - if (ret_value<0) + if (ret_value<0) HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "optimized write failed") } else { ret_value = (io_info->ops.read)(io_info, chunk_info->chunk_points,H5T_get_size(io_info->dset->shared->type), chunk_info->fspace,chunk_info->mspace,0, - buf); + buf); /* Check return value from optimized write */ - if (ret_value<0) + if (ret_value<0) HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "optimized read failed") } @@ -1241,7 +1241,7 @@ printf("inside independent IO mpi_rank = %d, chunk index = %d\n",mpi_rank,i); * * Purpose: Routine for the shared part of collective IO between multiple chunk collective IO and contiguous collective IO - + * * Return: Non-negative on success/Negative on failure * @@ -1251,16 +1251,16 @@ printf("inside independent IO mpi_rank = %d, chunk index = %d\n",mpi_rank,i); * *------------------------------------------------------------------------- */ -static herr_t +static herr_t H5D_inter_collective_io(H5D_io_info_t *io_info,const H5S_t *file_space,const H5S_t *mem_space, - haddr_t addr, const void *buf, hbool_t do_write ) + haddr_t addr, const void *buf, hbool_t do_write ) { size_t mpi_buf_count, mpi_file_count; /* Number of "objects" to transfer */ MPI_Datatype mpi_file_type,mpi_buf_type; hsize_t mpi_buf_offset, mpi_file_offset; /* Offset within dataset where selection (ie. MPI type) begins */ hbool_t mbt_is_derived=0, /* Whether the buffer (memory) type is derived and needs to be free'd */ - mft_is_derived=0; /* Whether the file type is derived and needs to be free'd */ + mft_is_derived=0; /* Whether the file type is derived and needs to be free'd */ H5D_common_coll_info_t coll_info; herr_t ret_value = SUCCEED; /* return value */ @@ -1274,11 +1274,11 @@ H5D_inter_collective_io(H5D_io_info_t *io_info,const H5S_t *file_space,const H5S if(H5S_mpio_space_type(mem_space,H5T_get_size(io_info->dset->shared->type), &mpi_buf_type,&mpi_buf_count,&mpi_buf_offset,&mbt_is_derived)<0) HGOTO_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"couldn't create MPI buffer type"); - + } else { /* For non-selection, participate with a none MPI derived datatype, the count is 0. */ - mpi_buf_type = MPI_BYTE; + mpi_buf_type = MPI_BYTE; mpi_file_type = MPI_BYTE; mpi_file_count = 0; mpi_buf_count = 0; @@ -1305,7 +1305,7 @@ printf("before leaving inter_collective_io ret_value = %d\n",ret_value); * Function: H5D_final_collective_io * * Purpose: Routine for the common part of collective IO with different storages. - + * * Return: Non-negative on success/Negative on failure * @@ -1315,13 +1315,13 @@ printf("before leaving inter_collective_io ret_value = %d\n",ret_value); * *------------------------------------------------------------------------- */ -static herr_t +static herr_t H5D_final_collective_io(H5D_io_info_t *io_info,MPI_Datatype*mpi_file_type,MPI_Datatype *mpi_buf_type, - H5D_common_coll_info_t* coll_info, const void *buf, hbool_t do_write) + H5D_common_coll_info_t* coll_info, const void *buf, hbool_t do_write) { - int mpi_code; /* MPI return code */ + int mpi_code; /* MPI return code */ hbool_t plist_is_setup=0; /* Whether the dxpl has been customized */ herr_t ret_value = SUCCEED; @@ -1338,7 +1338,7 @@ H5D_final_collective_io(H5D_io_info_t *io_info,MPI_Datatype*mpi_file_type,MPI_Da plist_is_setup=1; #ifdef KENT HDfprintf(stdout,"chunk addr %Hu\n",coll_info->chunk_addr); - printf("mpi_buf_count %d\n",coll_info->mpi_buf_count); + printf("mpi_buf_count %d\n",coll_info->mpi_buf_count); #endif if(do_write) { ret_value = (io_info->ops.write)(io_info, @@ -1348,7 +1348,7 @@ H5D_final_collective_io(H5D_io_info_t *io_info,MPI_Datatype*mpi_file_type,MPI_Da #ifdef KENT printf("ret_value after final collective IO= %d\n",ret_value); #endif - if (ret_value<0) + if (ret_value<0) HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "optimized write failed") } else { @@ -1356,11 +1356,11 @@ H5D_final_collective_io(H5D_io_info_t *io_info,MPI_Datatype*mpi_file_type,MPI_Da coll_info->mpi_buf_count,0,NULL,NULL,coll_info->chunk_addr, buf); /* Check return value from optimized write */ - if (ret_value<0) + if (ret_value<0) HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "optimized read failed") } done: - + /* Reset the dxpl settings */ if(plist_is_setup) { if(H5FD_mpi_teardown_collective(io_info->dxpl_id)<0) @@ -1371,7 +1371,7 @@ H5D_final_collective_io(H5D_io_info_t *io_info,MPI_Datatype*mpi_file_type,MPI_Da if (coll_info->mbt_is_derived) { if (MPI_SUCCESS != (mpi_code= MPI_Type_free( mpi_buf_type ))) HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code); - + } if (coll_info->mft_is_derived) { if (MPI_SUCCESS != (mpi_code= MPI_Type_free( mpi_file_type ))) @@ -1392,7 +1392,7 @@ H5D_final_collective_io(H5D_io_info_t *io_info,MPI_Datatype*mpi_file_type,MPI_Da Description: root will collective all chunk addresses and broadcast towards other processes. - + Parameters: Input: H5D_io_info_t* io_info, @@ -1430,7 +1430,7 @@ H5D_pre_sort_chunk(H5D_io_info_t *io_info,int total_chunks,haddr_t total_chunk_a if(MPI_SUCCESS != (mpi_code = MPI_Type_commit(&chunk_addrtype))) HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code); - + mpi_type_cleanup = 1; if(mpi_rank == root) { @@ -1441,7 +1441,7 @@ H5D_pre_sort_chunk(H5D_io_info_t *io_info,int total_chunks,haddr_t total_chunk_a /* Broadcasting the MPI_IO option info. and chunk address info. */ if(MPI_SUCCESS !=(mpi_code = MPI_Bcast(total_chunk_addr_array,1,chunk_addrtype,root,comm))) HMPI_GOTO_ERROR(FAIL, "MPI_BCast failed", mpi_code); - + done: if(mpi_type_cleanup){ @@ -1462,12 +1462,12 @@ done: For most cases, the chunk address has already been sorted in increasing order. The special sorting flag is used to optimize this common case. quick sort is used for necessary sorting. - + Parameters: Input: H5D_io_info_t* io_info, fm_map *fm(global chunk map struct) - Input/Output: H5D_chunk_addr_info_t chunk_addr_info_array[] : array to store chunk address and information - many_chunk_opt : flag to optimize the way to obtain chunk addresses + Input/Output: H5D_chunk_addr_info_t chunk_addr_info_array[] : array to store chunk address and information + many_chunk_opt : flag to optimize the way to obtain chunk addresses for many chunks * * Return: Non-negative on success/Negative on failure @@ -1479,7 +1479,7 @@ done: *------------------------------------------------------------------------- */ -static herr_t +static herr_t H5D_sort_chunk(H5D_io_info_t * io_info, fm_map *fm, H5D_chunk_addr_info_t chunk_addr_info_array[], @@ -1500,7 +1500,7 @@ H5D_sort_chunk(H5D_io_info_t * io_info, H5D_storage_t store; /*union of EFL and chunk pointer in file space */ hbool_t do_sort = FALSE; herr_t ret_value = SUCCEED; /*return value */ - + FUNC_ENTER_NOAPI_NOINIT(H5D_sort_chunk) num_chunks = H5SL_count(fm->fsel); @@ -1563,7 +1563,7 @@ printf("Coming inside H5D_OBTAIN_ALL_CHUNK_ADDR_COL\n"); if(HADDR_UNDEF==(chunk_addr = H5D_istore_get_addr(io_info,NULL))) HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL,"couldn't get chunk info from skipped list"); } - else + else chunk_addr = total_chunk_addr_array[chunk_info->index]; chunk_addr_info_array[i].chunk_addr = chunk_addr; chunk_addr_info_array[i].chunk_info = *chunk_info; @@ -1574,12 +1574,12 @@ printf("Coming inside H5D_OBTAIN_ALL_CHUNK_ADDR_COL\n"); chunk_info = H5SL_item(chunk_node); store.chunk.offset = chunk_info->coords; store.chunk.index = chunk_info->index; - + if(many_chunk_opt == H5D_OBTAIN_ONE_CHUNK_ADDR_IND){ if(HADDR_UNDEF==(chunk_addr = H5D_istore_get_addr(io_info,NULL))) HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL,"couldn't get chunk info from skipped list"); } - else + else chunk_addr = total_chunk_addr_array[chunk_info->index]; if(chunk_addr < chunk_addr_info_array[i].chunk_addr) do_sort = TRUE; @@ -1605,7 +1605,7 @@ done: } FUNC_LEAVE_NOAPI(ret_value) } /* end H5D_sort_chunk() */ - + /*------------------------------------------------------------------------- * Function: H5D_obtain_mpio_mode @@ -1616,11 +1616,11 @@ done: Description: 1) Each process provides two piece of information for all chunks with selection - a) chunk index + a) chunk index b) wheather this chunk is regular(for MPI derived datatype not working case) 2) Gather all the information to the root process - + 3) Root process will do the following: a) Obtain chunk address for all chunks in this data space b) With the consideration of the user option, calculate IO mode for each chunk @@ -1628,7 +1628,7 @@ done: in order to do MPI Bcast only once d) MPI Bcast the IO mode and chunk address information for each chunk. 4) Each process then retrieves IO mode and chunk address information to assign_io_mode and chunk_addr. - + Parameters: Input: H5D_io_info_t* io_info, @@ -1645,8 +1645,8 @@ done: *------------------------------------------------------------------------- */ -static herr_t -H5D_obtain_mpio_mode(H5D_io_info_t* io_info, +static herr_t +H5D_obtain_mpio_mode(H5D_io_info_t* io_info, fm_map *fm, uint8_t assign_io_mode[], haddr_t chunk_addr[]) @@ -1661,7 +1661,7 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, uint8_t* mergebuf=NULL; uint8_t* tempbuf; - H5SL_node_t* chunk_node; + H5SL_node_t* chunk_node; H5D_chunk_info_t* chunk_info; MPI_Datatype bastype[2]; @@ -1688,7 +1688,7 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, FUNC_ENTER_NOAPI_NOINIT(H5D_obtain_mpio_mode) /* Assign the rank 0 to the root */ - root = 0; + root = 0; comm = io_info->comm; /* Obtain the number of process and the current rank of the process */ @@ -1696,7 +1696,7 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, HGOTO_ERROR (H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank"); if((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file))<0) HGOTO_ERROR (H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size"); - + /* Allocate memory */ ori_total_chunks = fm->total_chunks; H5_ASSIGN_OVERFLOW(total_chunks,ori_total_chunks,hsize_t,int); @@ -1704,30 +1704,30 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, /* Obtain the data transfer properties */ if(NULL == (plist = H5I_object(io_info->dxpl_id))) HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a file access property list") - + percent_nproc_per_chunk=H5P_peek_unsigned(plist,H5D_XFER_MPIO_CHUNK_OPT_RATIO_NAME); #if defined(H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS) && defined(H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS) - + chunk_opt_mode=(H5FD_mpio_chunk_opt_t)H5P_peek_unsigned(plist,H5D_XFER_MPIO_CHUNK_OPT_HARD_NAME); if((chunk_opt_mode == H5FD_MPIO_CHUNK_MULTI_IO) || (percent_nproc_per_chunk == 0)){ if(H5D_istore_chunkmap(io_info,total_chunks,chunk_addr,fm->down_chunks)<0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get chunk address"); + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get chunk address"); for(ic = 0; ic<total_chunks;ic++) assign_io_mode[ic] = H5D_CHUNK_IO_MODE_COL; goto done; } -#endif +#endif threshold_nproc_per_chunk = mpi_size * percent_nproc_per_chunk/100; io_mode_info = (uint8_t *)H5MM_calloc(total_chunks*sizeof(MPI_BYTE)); mergebuf = H5MM_malloc((sizeof(haddr_t)+sizeof(MPI_BYTE))*total_chunks); tempbuf = mergebuf + sizeof(MPI_BYTE)*total_chunks; - if(mpi_rank == root) + if(mpi_rank == root) recv_io_mode_info = (uint8_t *)H5MM_malloc(total_chunks*sizeof(MPI_BYTE)*mpi_size); - + mem_cleanup = 1; chunk_node = H5SL_first(fm->fsel); @@ -1750,7 +1750,7 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, chunk_node = H5SL_next(chunk_node); } - + /*Create sent MPI derived datatype */ if(MPI_SUCCESS !=(mpi_code = MPI_Type_contiguous(total_chunks,MPI_BYTE,&stype))) HMPI_GOTO_ERROR(FAIL, "MPI_Comm_rank failed", mpi_code); @@ -1764,7 +1764,7 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, basdisp[0] = 0; basdisp[1] = (MPI_Aint)(sizeof(MPI_BYTE)*total_chunks);/* may need to check overflow */ bastype[0] = MPI_BYTE; - + if(MPI_SUCCESS !=(mpi_code = MPI_Type_contiguous(sizeof(haddr_t),MPI_BYTE,&chunk_addrtype))) HMPI_GOTO_ERROR(FAIL, "MPI_Type_contiguous failed", mpi_code); if(MPI_SUCCESS != (mpi_code = MPI_Type_commit(&chunk_addrtype))) @@ -1792,7 +1792,7 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, int* ind_this_chunk; #endif - /* pre-computing: calculate number of processes and + /* pre-computing: calculate number of processes and regularity of the selection occupied in each chunk */ nproc_per_chunk = (int*)H5MM_calloc(total_chunks*sizeof(int)); #if !defined(H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS) || !defined(H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS) @@ -1866,7 +1866,7 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, new_value = 0; if(H5Pset(io_info->dxpl_id,H5D_XFER_COLL_CHUNK_MULTI_RATIO_COLL_NAME,&new_value)<0) HGOTO_ERROR(H5E_PLIST, H5E_UNSUPPORTED, FAIL, "unable to set property value"); -#else +#else for(ic = 0; ic < total_chunks; ic++){ if(assign_io_mode[ic] == H5D_CHUNK_IO_MODE_COL) { new_value = 0; @@ -1893,7 +1893,7 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, } } #endif - + done: if(mpi_type_cleanup) { @@ -1910,7 +1910,7 @@ done: if(mem_cleanup){ HDfree(io_mode_info); HDfree(mergebuf); - if(mpi_rank == root) + if(mpi_rank == root) HDfree(recv_io_mode_info); } @@ -1923,7 +1923,7 @@ H5D_cmp_chunk_addr(const void *chunk_addr_info1, const void *chunk_addr_info2) haddr_t addr1, addr2; FUNC_ENTER_NOAPI_NOINIT(H5D_cmp_chunk_addr) - + addr1 = ((const H5D_chunk_addr_info_t *)chunk_addr_info1)->chunk_addr; addr2 = ((const H5D_chunk_addr_info_t *)chunk_addr_info2)->chunk_addr; |