diff options
Diffstat (limited to 'src')
| -rw-r--r-- | src/H5Dmpio.c | 477 | ||||
| -rw-r--r-- | src/H5config.h.in | 7 |
2 files changed, 135 insertions, 349 deletions
diff --git a/src/H5Dmpio.c b/src/H5Dmpio.c index e646a7b..1c69305 100644 --- a/src/H5Dmpio.c +++ b/src/H5Dmpio.c @@ -17,33 +17,33 @@ * Programmer: rky 980813 * KY 2005 revised the code and made the change to support and optimize * collective IO support. - * Purpose: Functions to read/write directly between app buffer and file. + * Purpose: Functions to read/write directly between app buffer and file. * - * Beware of the ifdef'ed print statements. - * I didn't make them portable. + * Beware of the ifdef'ed print statements. + * I didn't make them portable. */ /****************/ /* Module Setup */ /****************/ -#define H5D_PACKAGE /*suppress error about including H5Dpkg */ +#define H5D_PACKAGE /* suppress error about including H5Dpkg */ /***********/ /* Headers */ /***********/ -#include "H5private.h" /* Generic Functions */ -#include "H5Dpkg.h" /* Datasets */ -#include "H5Eprivate.h" /* Error handling */ -#include "H5Fprivate.h" /* File access */ -#include "H5FDprivate.h" /* File drivers */ -#include "H5Iprivate.h" /* IDs */ -#include "H5MMprivate.h" /* Memory management */ -#include "H5Oprivate.h" /* Object headers */ -#include "H5Pprivate.h" /* Property lists */ -#include "H5Sprivate.h" /* Dataspaces */ -#include "H5Vprivate.h" /* Vector */ +#include "H5private.h" /* Generic Functions */ +#include "H5Dpkg.h" /* Datasets */ +#include "H5Eprivate.h" /* Error handling */ +#include "H5Fprivate.h" /* File access */ +#include "H5FDprivate.h" /* File drivers */ +#include "H5Iprivate.h" /* IDs */ +#include "H5MMprivate.h" /* Memory management */ +#include "H5Oprivate.h" /* Object headers */ +#include "H5Pprivate.h" /* Property lists */ +#include "H5Sprivate.h" /* Dataspaces */ +#include "H5Vprivate.h" /* Vector */ #ifdef H5_HAVE_PARALLEL @@ -107,20 +107,16 @@ static herr_t H5D_multi_chunk_collective_io(H5D_io_info_t *io_info, H5P_genplist_t *dx_plist); static herr_t H5D_multi_chunk_collective_io_no_opt(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, H5D_chunk_map_t *fm, H5P_genplist_t *dx_plist); -#ifdef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS static herr_t H5D_link_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, H5D_chunk_map_t *fm, int sum_chunk); -#endif /* H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS */ static herr_t H5D_inter_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, const H5S_t *file_space, const H5S_t *mem_space); static herr_t H5D_final_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, MPI_Datatype *mpi_file_type, MPI_Datatype *mpi_buf_type); -#ifdef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS static herr_t H5D_sort_chunk(H5D_io_info_t *io_info, const H5D_chunk_map_t *fm, H5D_chunk_addr_info_t chunk_addr_info_array[], int many_chunk_opt); -#endif /* H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS */ static herr_t H5D_obtain_mpio_mode(H5D_io_info_t *io_info, H5D_chunk_map_t *fm, H5P_genplist_t *dx_plist, uint8_t assign_io_mode[], haddr_t chunk_addr[]); static herr_t H5D_ioinfo_xfer_mode(H5D_io_info_t *io_info, H5P_genplist_t *dx_plist, @@ -145,15 +141,15 @@ static herr_t H5D_mpio_get_sum_chunk(const H5D_io_info_t *io_info, /*------------------------------------------------------------------------- - * Function: H5D_mpio_opt_possible + * Function: H5D_mpio_opt_possible * - * Purpose: Checks if an direct I/O transfer is possible between memory and + * Purpose: Checks if an direct I/O transfer is possible between memory and * the file. * - * Return: Success: Non-negative: TRUE or FALSE - * Failure: Negative + * Return: Sauccess: Non-negative: TRUE or FALSE + * Failure: Negative * - * Programmer: Quincey Koziol + * Programmer: Quincey Koziol * Wednesday, April 3, 2002 * *------------------------------------------------------------------------- @@ -225,19 +221,8 @@ H5D_mpio_opt_possible(const H5D_io_info_t *io_info, const H5S_t *file_space, * storage. For contiguous storage, mem_space and file_space won't change * when it it is doing disk IO. For chunking storage, mem_space will * change for different chunks. So for chunking storage, whether we can - * use collective IO will defer until each chunk IO is reached. For - * contiguous storage, if we find MPI-IO cannot support complicated MPI - * derived data type and the shape of data space is not regular, we will - * set use_par_opt_io = FALSE. + * use collective IO will defer until each chunk IO is reached. */ -#ifndef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS - if(io_info->dset->shared->layout.type == H5D_CONTIGUOUS) - if((H5S_SELECT_IS_REGULAR(file_space) != TRUE) || - (H5S_SELECT_IS_REGULAR(mem_space) != TRUE)) { - local_opinion = FALSE; - goto broadcast; - } /* end if */ -#endif /* Don't allow collective operations if filters need to be applied */ if(io_info->dset->shared->layout.type == H5D_CHUNKED) { @@ -245,24 +230,6 @@ H5D_mpio_opt_possible(const H5D_io_info_t *io_info, const H5S_t *file_space, local_opinion = FALSE; goto broadcast; } /* end if */ - -/* If H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS and H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS - * are defined, the HDF5 library will do collective IO if the application - * asks for it. - * - * If H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS is not defined and one or more - * processes are not participating in the IO, then collective IO is not - * assured. The library will check each process for the number of chunks - * it involves. If any process involves zero chunks, the library will use - * independent IO mode instead. - */ -#ifndef H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS - /* Check the number of chunks to perform I/O on */ - if(0 == H5SL_count(fm->sel_chunks)) { - local_opinion = FALSE; - goto broadcast; - } /* end if */ -#endif /* H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS */ } /* end if */ broadcast: @@ -280,11 +247,11 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_mpio_select_read + * Function: H5D_mpio_select_read * - * Purpose: MPI-IO function to read directly from app buffer to file. + * Purpose: MPI-IO function to read directly from app buffer to file. * - * Return: non-negative on success, negative on failure. + * Return: non-negative on success, negative on failure. * * Programmer: * @@ -301,7 +268,7 @@ H5D_mpio_select_read(const H5D_io_info_t *io_info, const H5D_type_info_t UNUSED H5_CHECK_OVERFLOW(mpi_buf_count, hsize_t, size_t); if(H5F_block_read(io_info->dset->oloc.file, H5FD_MEM_DRAW, store_contig->dset_addr, (size_t)mpi_buf_count, io_info->dxpl_id, io_info->u.rbuf) < 0) - HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "can't finish collective parallel read") + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "can't finish collective parallel read") done: FUNC_LEAVE_NOAPI(ret_value) @@ -309,11 +276,11 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_mpio_select_write + * Function: H5D_mpio_select_write * - * Purpose: MPI-IO function to write directly from app buffer to file. + * Purpose: MPI-IO function to write directly from app buffer to file. * - * Return: non-negative on success, negative on failure. + * Return: non-negative on success, negative on failure. * * Programmer: * @@ -339,14 +306,14 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_ioinfo_xfer_mode + * Function: H5D_ioinfo_xfer_mode * - * Purpose: Switch to between collective & independent MPI I/O + * Purpose: Switch to between collective & independent MPI I/O * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Quincey Koziol - * Friday, August 12, 2005 + * Programmer: Quincey Koziol + * Friday, August 12, 2005 * *------------------------------------------------------------------------- */ @@ -354,7 +321,7 @@ static herr_t H5D_ioinfo_xfer_mode(H5D_io_info_t *io_info, H5P_genplist_t *dx_plist, H5FD_mpio_xfer_t xfer_mode) { - herr_t ret_value = SUCCEED; /*return value */ + herr_t ret_value = SUCCEED; /* return value */ FUNC_ENTER_NOAPI_NOINIT(H5D_ioinfo_xfer_mode) @@ -383,15 +350,15 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_ioinfo_coll_opt_mode + * Function: H5D_ioinfo_coll_opt_mode * - * Purpose: Switch between using collective & independent MPI I/O w/file + * Purpose: Switch between using collective & independent MPI I/O w/file * set view * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: MuQun Yang - * Oct. 5th, 2006 + * Programmer: MuQun Yang + * Oct. 5th, 2006 * *------------------------------------------------------------------------- */ @@ -399,7 +366,7 @@ static herr_t H5D_ioinfo_coll_opt_mode(H5D_io_info_t *io_info, H5P_genplist_t *dx_plist, H5FD_mpio_collective_opt_t coll_opt_mode) { - herr_t ret_value = SUCCEED; /*return value */ + herr_t ret_value = SUCCEED; /* return value */ FUNC_ENTER_NOAPI_NOINIT(H5D_ioinfo_coll_opt_mode) @@ -414,14 +381,14 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_mpio_get_min_chunk + * Function: H5D_mpio_get_min_chunk * - * Purpose: Routine for obtaining minimum number of chunks to cover + * Purpose: Routine for obtaining minimum number of chunks to cover * hyperslab selection selected by all processors. * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Muqun Yang + * Programmer: Muqun Yang * Monday, Feb. 13th, 2006 * *------------------------------------------------------------------------- @@ -449,14 +416,14 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_mpio_get_sum_chunk + * Function: H5D_mpio_get_sum_chunk * - * Purpose: Routine for obtaining total number of chunks to cover + * Purpose: Routine for obtaining total number of chunks to cover * hyperslab selection selected by all processors. * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Muqun Yang + * Programmer: Muqun Yang * Monday, Feb. 13th, 2006 * *------------------------------------------------------------------------- @@ -487,14 +454,14 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_contig_collective_read + * Function: H5D_contig_collective_read * - * Purpose: Reads directly from contiguous data in file into application + * Purpose: Reads directly from contiguous data in file into application * memory using collective I/O. * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Quincey Koziol + * Programmer: Quincey Koziol * Tuesday, March 4, 2008 * *------------------------------------------------------------------------- @@ -514,7 +481,7 @@ H5D_contig_collective_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_i /* Call generic internal collective I/O routine */ if(H5D_inter_collective_io(io_info, type_info, file_space, mem_space) < 0) - HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "couldn't finish shared collective MPI-IO") + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "couldn't finish shared collective MPI-IO") done: FUNC_LEAVE_NOAPI(ret_value) @@ -522,14 +489,14 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_contig_collective_write + * Function: H5D_contig_collective_write * - * Purpose: Write directly to contiguous data in file from application + * Purpose: Write directly to contiguous data in file from application * memory using collective I/O. * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Quincey Koziol + * Programmer: Quincey Koziol * Tuesday, March 4, 2008 * *------------------------------------------------------------------------- @@ -549,7 +516,7 @@ H5D_contig_collective_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_ /* Call generic internal collective I/O routine */ if(H5D_inter_collective_io(io_info, type_info, file_space, mem_space) < 0) - HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "couldn't finish shared collective MPI-IO") + HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "couldn't finish shared collective MPI-IO") done: FUNC_LEAVE_NOAPI(ret_value) @@ -557,9 +524,9 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_chunk_collective_io + * Function: H5D_chunk_collective_io * - * Purpose: Routine for + * 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 @@ -570,7 +537,7 @@ done: * 3. Build up the final MPI derived datatype * 4. Set up collective IO property list * 5. Do IO - * For option b) + * For option b) * 1. Use MPI_gather and MPI_Bcast to obtain information of *collective/independent/none* * IO mode for each chunk of the selection * 2. Depending on whether the IO mode is collective or independent or none, @@ -579,9 +546,9 @@ done: * 3. Set up collective IO property list for collective mode * 4. DO IO * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Muqun Yang + * Programmer: Muqun Yang * Monday, Feb. 13th, 2006 * *------------------------------------------------------------------------- @@ -636,43 +603,9 @@ H5D_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info else temp_not_link_io = TRUE; #endif - } /* end else */ - -#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. */ -#ifdef H5_HAVE_INSTRUMENTED_LIBRARY - { int new_value; - htri_t check_prop; - check_prop = H5Pexist(io_info->dxpl_id, H5D_XFER_COLL_CHUNK_LINK_TO_MULTI); - if(check_prop > 0) { - new_value = 1; - if(H5Pset(io_info->dxpl_id, H5D_XFER_COLL_CHUNK_LINK_TO_MULTI, &new_value) < 0) - HGOTO_ERROR(H5E_IO, H5E_CANTSET, FAIL, "unable to set property value") - - } - }/* add this property because the library changes the option from one link to multiple chunks.*/ -#endif - } - if(io_option == H5D_ONE_LINK_CHUNK_IO_MORE_OPT){ - io_option = H5D_MULTI_CHUNK_IO_MORE_OPT; -#ifdef H5_HAVE_INSTRUMENTED_LIBRARY - { int new_value; - htri_t check_prop; - check_prop = H5Pexist(io_info->dxpl_id, H5D_XFER_COLL_CHUNK_LINK_TO_MULTI_OPT); - if(check_prop > 0) { - new_value = 1; - if(H5Pset(io_info->dxpl_id, H5D_XFER_COLL_CHUNK_LINK_TO_MULTI_OPT, &new_value) < 0) - HGOTO_ERROR(H5E_IO, H5E_CANTSET, FAIL, "unable to set property value") - - } - }/* add this property because the library changes the option from one link to multiple chunks.*/ -#endif - } -#endif + } /* end else */ #ifdef H5_HAVE_INSTRUMENTED_LIBRARY -{ htri_t check_prop; int new_value; @@ -709,17 +642,14 @@ H5D_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info HGOTO_ERROR(H5E_IO, H5E_CANTSET, FAIL, "unable to set property value") } /* end if */ } /* end if */ -} #endif /* step 2: Go ahead to do IO.*/ -#ifdef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS if(io_option == H5D_ONE_LINK_CHUNK_IO || io_option == H5D_ONE_LINK_CHUNK_IO_MORE_OPT) { if(H5D_link_chunk_collective_io(io_info, type_info, fm, sum_chunk) < 0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish linked chunk MPI-IO") } /* end if */ else -#endif /* H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS */ if(io_option == H5D_MULTI_CHUNK_IO) { if(H5D_multi_chunk_collective_io_no_opt(io_info, type_info, fm, dx_plist) < 0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish multiple chunk MPI-IO") @@ -735,14 +665,14 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_chunk_collective_read + * Function: H5D_chunk_collective_read * - * Purpose: Reads directly from chunks in file into application memory + * Purpose: Reads directly from chunks in file into application memory * using collective I/O. * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Quincey Koziol + * Programmer: Quincey Koziol * Tuesday, March 4, 2008 * *------------------------------------------------------------------------- @@ -766,14 +696,14 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_chunk_collective_write + * Function: H5D_chunk_collective_write * - * Purpose: Write directly to chunks in file from application memory + * Purpose: Write directly to chunks in file from application memory * using collective I/O. * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Quincey Koziol + * Programmer: Quincey Koziol * Tuesday, March 4, 2008 * *------------------------------------------------------------------------- @@ -795,21 +725,20 @@ done: FUNC_LEAVE_NOAPI(ret_value) } /* end H5D_chunk_collective_write() */ -#ifdef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS /*------------------------------------------------------------------------- - * Function: H5D_link_chunk_collective_io + * Function: H5D_link_chunk_collective_io * - * Purpose: Routine for one collective IO with one MPI derived datatype to link with all chunks + * Purpose: Routine for one collective IO with one MPI derived datatype to link with all chunks * * 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 * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Muqun Yang + * Programmer: Muqun Yang * Monday, Feb. 13th, 2006 * *------------------------------------------------------------------------- @@ -833,7 +762,7 @@ H5D_link_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type hbool_t *chunk_mft_is_derived_array = NULL; /* Flags to indicate each chunk's MPI file datatype is derived */ hbool_t *chunk_mbt_is_derived_array = NULL; /* Flags to indicate each chunk's MPI memory datatype is derived */ int *chunk_mpi_file_counts = NULL; /* Count of MPI file datatype for each chunk */ - int *chunk_mpi_mem_counts = NULL; /* Count of MPI memory datatype for each chunk */ + int *chunk_mpi_mem_counts = NULL; /* Count of MPI memory datatype for each chunk */ int mpi_code; /* MPI return code */ herr_t ret_value = SUCCEED; @@ -871,21 +800,22 @@ H5D_link_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type /* Check for this process having selection in this chunk */ chunk_node = H5SL_first(fm->sel_chunks); - if(chunk_node == NULL) { - /* Set the dataspace info for I/O to NULL, this process doesn't have any I/O to perform */ - fspace = mspace = NULL; - } /* end if */ - else { - H5D_chunk_info_t *chunk_info; - - /* Get the chunk info, for the selection in the chunk */ - if(NULL == (chunk_info = H5SL_item(chunk_node))) - HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL, "couldn't get chunk info from skipped list") - /* Set the dataspace info for I/O */ - fspace = chunk_info->fspace; - mspace = chunk_info->mspace; - } /* end else */ + if(chunk_node == NULL) { + /* Set the dataspace info for I/O to NULL, this process doesn't have any I/O to perform */ + fspace = mspace = NULL; + } /* end if */ + else { + H5D_chunk_info_t *chunk_info; + + /* Get the chunk info, for the selection in the chunk */ + if(NULL == (chunk_info = H5SL_item(chunk_node))) + HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL, "couldn't get chunk info from skipped list") + + /* Set the dataspace info for I/O */ + fspace = chunk_info->fspace; + mspace = chunk_info->mspace; + } /* end else */ /* Set up the base storage address for this chunk */ io_info->store = &ctg_store; @@ -943,6 +873,7 @@ if(H5DEBUG(D)) if(H5D_sort_chunk(io_info, fm, chunk_addr_info_array, sum_chunk) < 0) HGOTO_ERROR(H5E_DATASPACE, H5E_CANTSWAP, FAIL, "unable to sort chunk address") ctg_store.contig.dset_addr = chunk_addr_info_array[0].chunk_addr; + #ifdef H5D_DEBUG if(H5DEBUG(D)) HDfprintf(H5DEBUG(D),"after sorting the chunk address \n"); @@ -1055,7 +986,7 @@ if(H5DEBUG(D)) if(chunk_mpi_file_counts) H5MM_xfree(chunk_mpi_file_counts); if(chunk_mbt_is_derived_array) - H5MM_xfree(chunk_mbt_is_derived_array); + H5MM_xfree(chunk_mbt_is_derived_array); if(chunk_mft_is_derived_array) H5MM_xfree(chunk_mft_is_derived_array); @@ -1067,22 +998,21 @@ if(H5DEBUG(D)) FUNC_LEAVE_NOAPI(ret_value) } /* end H5D_link_chunk_collective_io */ -#endif /* H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS */ /*------------------------------------------------------------------------- - * Function: H5D_multi_chunk_collective_io + * Function: H5D_multi_chunk_collective_io * - * Purpose: To do IO per chunk according to IO mode(collective/independent/none) + * Purpose: To do IO per chunk according to IO mode(collective/independent/none) * * 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 * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Muqun Yang + * Programmer: Muqun Yang * Monday, Feb. 13th, 2006 * *------------------------------------------------------------------------- @@ -1169,7 +1099,7 @@ if(H5DEBUG(D)) /* Pass in chunk's coordinates in a union. */ store.chunk.offset = chunk_info->coords; store.chunk.index = chunk_info->index; - } /* end if */ + } /* end if */ /* Collective IO for this chunk, * Note: even there is no selection for this process, the process still @@ -1185,10 +1115,10 @@ if(H5DEBUG(D)) if(chunk_info) { fspace = chunk_info->fspace; mspace = chunk_info->mspace; - } /* end if */ - else { + } /* end if */ + else { fspace = mspace = NULL; - } /* end else */ + } /* end else */ /* Switch back to collective I/O */ if(last_xfer_mode != H5FD_MPIO_COLLECTIVE) { @@ -1208,8 +1138,8 @@ if(H5DEBUG(D)) /* Perform the I/O */ if(H5D_inter_collective_io(&ctg_io_info, type_info, fspace, mspace) < 0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish shared collective MPI-IO") - } /* end if */ - else { /* possible independent IO for this chunk */ + } /* end if */ + else { /* possible independent IO for this chunk */ #ifdef H5D_DEBUG if(H5DEBUG(D)) HDfprintf(H5DEBUG(D),"inside independent IO mpi_rank = %d, chunk index = %Zu\n", mpi_rank, u); @@ -1217,81 +1147,6 @@ if(H5DEBUG(D)) HDassert(chunk_io_option[u] == 0); -#if !defined(H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS) || !defined(H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS) - /* Check if this process has something to do with this chunk */ - if(chunk_info) { - H5D_io_info_t *chk_io_info; /* Pointer to I/O info object for this chunk */ - H5D_chunk_ud_t udata; /* B-tree pass-through */ - void *chunk; /* Pointer to the data chunk in cache */ - uint32_t accessed_bytes; /* Total accessed size in a chunk */ - htri_t cacheable; /* Whether the chunk is cacheable */ - - /* Switch to independent I/O */ - if(last_xfer_mode != H5FD_MPIO_INDEPENDENT) { - if(H5D_ioinfo_xfer_mode(io_info, dx_plist, H5FD_MPIO_INDEPENDENT) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't switch to independent I/O") - last_xfer_mode = H5FD_MPIO_INDEPENDENT; - } /* end if */ - - /* Load the chunk into cache. But if the whole chunk is written, - * simply allocate space instead of load the chunk. - */ - if(H5D_chunk_lookup(io_info->dset, io_info->dxpl_id, - chunk_info->coords, chunk_info->index, &udata) < 0) - HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL, "couldn't get chunk info from skipped list") - - /* Load the chunk into cache and lock it. */ - if((cacheable = H5D_chunk_cacheable(io_info, udata.addr, - io_info->op_type == H5D_IO_OP_WRITE)) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't tell if chunk is cacheable") - if(cacheable) { - hbool_t entire_chunk = TRUE; /* Whether whole chunk is selected */ - - /* Compute # of bytes accessed in chunk */ - accessed_bytes = chunk_info->chunk_points * type_info->src_type_size; - - /* Determine if we will access all the data in the chunk */ - if(((io_info->op_type == H5D_IO_OP_WRITE) && (accessed_bytes != ctg_store.contig.dset_size)) - || (io_info->op_type != H5D_IO_OP_WRITE)) - entire_chunk = FALSE; - - /* Lock the chunk into the cache */ - if(NULL == (chunk = H5D_chunk_lock(io_info, &udata, entire_chunk))) - HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk") - - /* Set up the storage buffer information for this chunk */ - cpt_store.compact.buf = chunk; - - /* Point I/O info at contiguous I/O info for this chunk */ - chk_io_info = &cpt_io_info; - } /* end if */ - else { - /* Set up the storage address information for this chunk */ - ctg_store.contig.dset_addr = udata.addr; - - /* No chunk cached */ - chunk = NULL; - - /* Point I/O info at temporary I/O info for this chunk */ - chk_io_info = &ctg_io_info; - } /* end else */ - - if(io_info->op_type == H5D_IO_OP_WRITE) { - if((io_info->io_ops.single_write)(chk_io_info, type_info, - (hsize_t)chunk_info->chunk_points, chunk_info->fspace, chunk_info->mspace) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "optimized write failed") - } /* end if */ - else { - if((io_info->io_ops.single_read)(chk_io_info, type_info, - (hsize_t)chunk_info->chunk_points, chunk_info->fspace, chunk_info->mspace) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "optimized read failed") - } /* end else */ - - /* Release the cache lock on the chunk. */ - if(chunk && H5D_chunk_unlock(io_info, &udata, (io_info->op_type == H5D_IO_OP_WRITE), chunk, accessed_bytes) < 0) - HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to unlock raw data chunk") - } /* end if */ -#else /* !defined(H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS) || !defined(H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS) */ /* Set the file & memory dataspaces */ if(chunk_info) { fspace = chunk_info->fspace; @@ -1318,7 +1173,6 @@ if(H5DEBUG(D)) if(H5DEBUG(D)) HDfprintf(H5DEBUG(D),"after inter collective IO\n"); #endif -#endif /* !defined(H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS) || !defined(H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS) */ } /* end else */ } /* end for */ @@ -1333,9 +1187,9 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_multi_chunk_collective_io_no_opt + * Function: H5D_multi_chunk_collective_io_no_opt * - * Purpose: To do collective IO without any optimization per chunk base + * Purpose: To do collective IO without any optimization per chunk base * The internal independent IO inside HDF5 cannot handle * non-contiguous(or with holes) storage efficiently. * Under this case, the one independent IO call may consist of @@ -1352,9 +1206,9 @@ done: * The HDF5 library won't do any IO management but leave it to MPI-IO to figure * out. * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Muqun Yang + * Programmer: Muqun Yang * Monday, Feb. 13th, 2006 * *------------------------------------------------------------------------- @@ -1416,7 +1270,7 @@ if(H5DEBUG(D)) { /* Iterate through chunks to be operated on */ while(chunk_node) { H5D_chunk_info_t *chunk_info; /* chunk information */ - H5D_chunk_ud_t udata; /* B-tree pass-through */ + H5D_chunk_ud_t udata; /* B-tree pass-through */ hbool_t make_ind, make_coll; /* Flags to indicate that the MPI mode should change */ /* Get the actual chunk information from the skip list node */ @@ -1437,20 +1291,6 @@ if(H5DEBUG(D)) { if(count_chunk > min_chunk) /* Switch to independent I/O (permanently) */ make_ind = TRUE; -#ifndef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS -/* This case needs to be improved to check if the selected space - is regular. If all selections are regular, collective IO can still be done. - However, since we find an MPI-IO bug at a DOE machine(mcr) that cannot - handle collective I/O selection for this case correctly, - we turn off this optimization but leave the following code - for future optimization. Otherwise, the following else {} doesn't make sense. - KY 2006/8/4/ */ - else { - /* Switch to independent I/O (temporarily) */ - make_ind = TRUE; - make_coll = TRUE; - } /* end else */ -#endif /* H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS */ /* Retrieve the chunk's address */ if(H5D_chunk_lookup(io_info->dset, io_info->dxpl_id, chunk_info->coords, @@ -1542,14 +1382,14 @@ done: /*------------------------------------------------------------------------- - * Function: H5D_inter_collective_io + * Function: H5D_inter_collective_io * - * Purpose: Routine for the shared part of collective IO between multiple chunk + * 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 + * Return: Non-negative on success/Negative on failure * - * Programmer: Muqun Yang + * Programmer: Muqun Yang * Monday, Feb. 13th, 2006 * *------------------------------------------------------------------------- @@ -1611,13 +1451,13 @@ if(H5DEBUG(D)) /*------------------------------------------------------------------------- - * Function: H5D_final_collective_io + * Function: H5D_final_collective_io * - * Purpose: Routine for the common part of collective IO with different storages. + * Purpose: Routine for the common part of collective IO with different storages. * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Muqun Yang + * Programmer: Muqun Yang * Monday, Feb. 13th, 2006 * *------------------------------------------------------------------------- @@ -1626,7 +1466,7 @@ static herr_t H5D_final_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t mpi_buf_count, MPI_Datatype *mpi_file_type, MPI_Datatype *mpi_buf_type) { - hbool_t plist_is_setup = FALSE; /* Whether the dxpl has been customized */ + hbool_t plist_is_setup = FALSE; /* Whether the dxpl has been customized */ herr_t ret_value = SUCCEED; FUNC_ENTER_NOAPI_NOINIT(H5D_final_collective_io) @@ -1638,7 +1478,7 @@ H5D_final_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info if(io_info->op_type == H5D_IO_OP_WRITE) { if((io_info->io_ops.single_write)(io_info, type_info, mpi_buf_count, NULL, NULL) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "optimized write failed") + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "optimized write failed") } /* end if */ else { if((io_info->io_ops.single_read)(io_info, type_info, mpi_buf_count, NULL, NULL) < 0) @@ -1658,12 +1498,11 @@ if(H5DEBUG(D)) FUNC_LEAVE_NOAPI(ret_value) } /* end H5D_final_collective_io */ -#ifdef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS /*------------------------------------------------------------------------- - * Function: H5D_sort_chunk + * Function: H5D_sort_chunk * - * Purpose: Routine to sort chunks in increasing order of chunk address + * Purpose: Routine to sort chunks in increasing order of chunk address * Each chunk address is also obtained. * * Description: @@ -1678,9 +1517,9 @@ if(H5DEBUG(D)) * many_chunk_opt : flag to optimize the way to obtain chunk addresses * for many chunks * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Muqun Yang + * Programmer: Muqun Yang * Monday, Feb. 13th, 2006 * *------------------------------------------------------------------------- @@ -1699,7 +1538,7 @@ H5D_sort_chunk(H5D_io_info_t *io_info, const H5D_chunk_map_t *fm, int mpi_size; /* Number of MPI processes */ int mpi_code; /* MPI return code */ int i; /* Local index variable */ - herr_t ret_value = SUCCEED; /* Return value */ + herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_NOAPI_NOINIT(H5D_sort_chunk) @@ -1740,14 +1579,15 @@ if(H5DEBUG(D)) /* Retrieve all the chunk addresses with process 0 */ if((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0) HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank") - if(mpi_rank == 0) { + + if(mpi_rank == 0) { if(H5D_chunk_addrmap(io_info, total_chunk_addr_array) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get chunk address") - } /* end if */ + } /* end if */ - /* Broadcasting the MPI_IO option info. and chunk address info. */ - if(MPI_SUCCESS != (mpi_code = MPI_Bcast(total_chunk_addr_array, (int)(sizeof(haddr_t) * fm->layout->u.chunk.nchunks), MPI_BYTE, (int)0, io_info->comm))) - HMPI_GOTO_ERROR(FAIL, "MPI_BCast failed", mpi_code) + /* Broadcasting the MPI_IO option info. and chunk address info. */ + if(MPI_SUCCESS != (mpi_code = MPI_Bcast(total_chunk_addr_array, (int)(sizeof(haddr_t) * fm->layout->u.chunk.nchunks), MPI_BYTE, (int)0, io_info->comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_BCast failed", mpi_code) } /* end if */ /* Start at first node in chunk skip list */ @@ -1801,13 +1641,12 @@ done: FUNC_LEAVE_NOAPI(ret_value) } /* end H5D_sort_chunk() */ -#endif /* H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS */ /*------------------------------------------------------------------------- - * Function: H5D_obtain_mpio_mode + * Function: H5D_obtain_mpio_mode * - * Purpose: Routine to obtain each io mode(collective,independent or none) for each chunk; + * Purpose: Routine to obtain each io mode(collective,independent or none) for each chunk; * Each chunk address is also obtained. * * Description: @@ -1833,9 +1672,9 @@ done: * Output: uint8_t assign_io_mode[], : IO mode, collective, independent or none * haddr_t chunk_addr[], : chunk address array for each chunk * - * Return: Non-negative on success/Negative on failure + * Return: Non-negative on success/Negative on failure * - * Programmer: Muqun Yang + * Programmer: Muqun Yang * Monday, Feb. 13th, 2006 * *------------------------------------------------------------------------- @@ -1846,9 +1685,7 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, H5D_chunk_map_t *fm, { int total_chunks; unsigned percent_nproc_per_chunk, threshold_nproc_per_chunk; -#if defined(H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS) && defined(H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS) H5FD_mpio_chunk_opt_t chunk_opt_mode; -#endif uint8_t* io_mode_info = NULL; uint8_t* recv_io_mode_info = NULL; uint8_t* mergebuf = NULL; @@ -1881,7 +1718,6 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, H5D_chunk_map_t *fm, /* Setup parameters */ H5_ASSIGN_OVERFLOW(total_chunks, fm->layout->u.chunk.nchunks, hsize_t, int); percent_nproc_per_chunk = H5P_peek_unsigned(dx_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(dx_plist, H5D_XFER_MPIO_CHUNK_OPT_HARD_NAME); if((chunk_opt_mode == H5FD_MPIO_CHUNK_MULTI_IO) || (percent_nproc_per_chunk == 0)) { if(H5D_chunk_addrmap(io_info, chunk_addr) < 0) @@ -1891,7 +1727,6 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, H5D_chunk_map_t *fm, HGOTO_DONE(SUCCEED) } /* end if */ -#endif threshold_nproc_per_chunk = mpi_size * percent_nproc_per_chunk/100; /* Allocate memory */ @@ -1907,16 +1742,7 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, H5D_chunk_map_t *fm, while(chunk_node) { chunk_info = H5SL_item(chunk_node); -#ifndef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS - /* regularity information: 1, selection information: 2 */ - if(H5S_SELECT_IS_REGULAR(chunk_info->fspace) == TRUE && - H5S_SELECT_IS_REGULAR(chunk_info->mspace) == TRUE) -#endif - io_mode_info[chunk_info->index] = H5D_CHUNK_SELECT_REG; /* this chunk is selected and is "regular" without defining H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS. */ -#ifndef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS - else - io_mode_info[chunk_info->index] = H5D_CHUNK_SELECT_IRREG; /* this chunk is selected and is irregular*/ -#endif + io_mode_info[chunk_info->index] = H5D_CHUNK_SELECT_REG; /* this chunk is selected and is "regular" */ chunk_node = H5SL_next(chunk_node); } /* end while */ @@ -1928,23 +1754,14 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, H5D_chunk_map_t *fm, if(mpi_rank == root) { int nproc; int* nproc_per_chunk; -#if !defined(H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS) || !defined(H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS) - int* ind_this_chunk; -#endif /* 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) - ind_this_chunk = (int*)H5MM_calloc(total_chunks * sizeof(int)); -#endif /* calculating the chunk address */ if(H5D_chunk_addrmap(io_info, chunk_addr) < 0) { HDfree(nproc_per_chunk); -#if !defined(H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS) || !defined(H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS) - HDfree(ind_this_chunk); -#endif HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get chunk address") } /* end if */ @@ -1956,29 +1773,14 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, H5D_chunk_map_t *fm, for(ic = 0; ic < total_chunks; ic++, tmp_recv_io_mode_info++) { if(*tmp_recv_io_mode_info != 0) { nproc_per_chunk[ic]++; -#ifndef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS - if(*tmp_recv_io_mode_info == H5D_CHUNK_SELECT_IRREG) - ind_this_chunk[ic] = 1; -#endif } /* end if */ -#ifndef H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS - else { - /*checking whether we have a selection in this chunk */ - ind_this_chunk[ic] = 1; - } /* end else */ -#endif } /* end for */ } /* end for */ /* Calculating MPIO mode for each chunk (collective, independent, none) */ for(ic = 0; ic < total_chunks; ic++) { if(nproc_per_chunk[ic] > MAX(1, threshold_nproc_per_chunk)) { -#if !defined(H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS) || !defined(H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS) - if(!ind_this_chunk[ic]) - assign_io_mode[ic] = H5D_CHUNK_IO_MODE_COL; -#else assign_io_mode[ic] = H5D_CHUNK_IO_MODE_COL; -#endif } /* end if */ } /* end for */ @@ -1988,9 +1790,6 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, H5D_chunk_map_t *fm, HDmemcpy(tempbuf, chunk_addr, sizeof(haddr_t) * total_chunks); HDfree(nproc_per_chunk); -#if !defined(H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS) || !defined(H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS) - HDfree(ind_this_chunk); -#endif } /* end if */ /* Broadcasting the MPI_IO option info. and chunk address info. */ @@ -2003,11 +1802,6 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, H5D_chunk_map_t *fm, #ifdef H5_HAVE_INSTRUMENTED_LIBRARY check_prop = H5Pexist(io_info->dxpl_id, H5D_XFER_COLL_CHUNK_MULTI_RATIO_COLL_NAME); if(check_prop > 0) { -#if !defined(H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS) || !defined(H5_MPI_SPECIAL_COLLECTIVE_IO_WORKS) - 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 for(ic = 0; ic < total_chunks; ic++) { if(assign_io_mode[ic] == H5D_CHUNK_IO_MODE_COL) { new_value = 0; @@ -2016,7 +1810,6 @@ H5D_obtain_mpio_mode(H5D_io_info_t* io_info, H5D_chunk_map_t *fm, break; } /* end if */ } /* end for */ -#endif } /* end if */ check_prop = H5Pexist(io_info->dxpl_id, H5D_XFER_COLL_CHUNK_MULTI_RATIO_IND_NAME); diff --git a/src/H5config.h.in b/src/H5config.h.in index 628996a..487ac1d 100644 --- a/src/H5config.h.in +++ b/src/H5config.h.in @@ -464,17 +464,10 @@ /* Define if the metadata trace file code is to be compiled in */ #undef METADATA_TRACE_FILE -/* Define if your system can handle complicated MPI derived datatype - correctly. */ -#undef MPI_COMPLEX_DERIVED_DATATYPE_WORKS - /* Define if your system's `MPI_File_set_size' function works for files over 2GB. */ #undef MPI_FILE_SET_SIZE_BIG -/* Define if your system can handle special collective IO properly. */ -#undef MPI_SPECIAL_COLLECTIVE_IO_WORKS - /* Define if we can violate pointer alignment restrictions */ #undef NO_ALIGNMENT_RESTRICTIONS |