From fed17ed3838d2cf73f8848c9d340a9139c0c02dc Mon Sep 17 00:00:00 2001 From: Quincey Koziol Date: Sat, 5 Jan 2019 22:31:42 -0600 Subject: HDFFV-10625 -- Implemented a process-0 read and then broadcast for collective read of full (HS_ALL), contiguous, atomic datasets by all the processes in the file communicator. --- src/H5CX.c | 92 ++++++++++++ src/H5CXprivate.h | 3 + src/H5Dio.c | 18 +-- src/H5Dmpio.c | 115 +++++++++++---- src/H5Dprivate.h | 8 +- src/H5FDmpio.c | 43 +++++- src/H5Pdxpl.c | 3 +- src/H5Ppublic.h | 3 +- src/H5T.c | 106 ++++++++++++++ src/H5Tprivate.h | 1 + testpar/t_bigio.c | 4 +- testpar/t_dset.c | 2 +- testpar/t_mdset.c | 173 +++++++++++++++------- testpar/t_pread.c | 422 +++++++++++++++++++++++++++++++++++++++++++++++++++--- 14 files changed, 869 insertions(+), 124 deletions(-) diff --git a/src/H5CX.c b/src/H5CX.c index 1f91ee2..0d20132 100644 --- a/src/H5CX.c +++ b/src/H5CX.c @@ -198,6 +198,7 @@ typedef struct H5CX_t { MPI_Datatype btype; /* MPI datatype for buffer, when using collective I/O */ MPI_Datatype ftype; /* MPI datatype for file, when using collective I/O */ hbool_t mpi_file_flushing; /* Whether an MPI-opened file is being flushed */ + hbool_t rank0_bcast; /* Whether a dataset meets read-with-rank0-and-bcast requirements */ #endif /* H5_HAVE_PARALLEL */ /* Cached DXPL properties */ @@ -261,6 +262,8 @@ typedef struct H5CX_t { hbool_t mpio_coll_chunk_multi_ratio_coll_set; /* Whether instrumented "collective chunk multi ratio coll" value is set */ int mpio_coll_chunk_multi_ratio_ind; /* Instrumented "collective chunk multi ratio ind" value (H5D_XFER_COLL_CHUNK_MULTI_RATIO_IND_NAME) */ hbool_t mpio_coll_chunk_multi_ratio_ind_set; /* Whether instrumented "collective chunk multi ratio ind" value is set */ + hbool_t mpio_coll_rank0_bcast; /* Instrumented "collective chunk multi ratio ind" value (H5D_XFER_COLL_CHUNK_MULTI_RATIO_IND_NAME) */ + hbool_t mpio_coll_rank0_bcast_set; /* Whether instrumented "collective chunk multi ratio ind" value is set */ #endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ #endif /* H5_HAVE_PARALLEL */ @@ -1254,6 +1257,32 @@ H5CX_get_mpi_file_flushing(void) FUNC_LEAVE_NOAPI((*head)->ctx.mpi_file_flushing) } /* end H5CX_get_mpi_file_flushing() */ + + +/*------------------------------------------------------------------------- + * Function: H5CX_get_mpio_rank0_bcast + * + * Purpose: Retrieves if the dataset meets read-with-rank0-and-bcast requirements for the current API call context. + * + * Return: Non-negative on success / Negative on failure + * + * Programmer: M. Breitenfeld + * December 31, 2018 + * + *------------------------------------------------------------------------- + */ +hbool_t +H5CX_get_mpio_rank0_bcast(void) +{ + H5CX_node_t **head = H5CX_get_my_context(); /* Get the pointer to the head of the API context, for this thread */ + + FUNC_ENTER_NOAPI_NOINIT_NOERR + + /* Sanity check */ + HDassert(head && *head); + + FUNC_LEAVE_NOAPI((*head)->ctx.rank0_bcast) +} /* end H5CX_get_mpio_rank0_bcast() */ #endif /* H5_HAVE_PARALLEL */ @@ -2185,6 +2214,34 @@ H5CX_set_mpi_file_flushing(hbool_t flushing) FUNC_LEAVE_NOAPI_VOID } /* end H5CX_set_mpi_file_flushing() */ + + +/*------------------------------------------------------------------------- + * Function: H5CX_set_mpio_rank0_bcast + * + * Purpose: Sets the "dataset meets read-with-rank0-and-bcast requirements" flag for the current API call context. + * + * Return: + * + * Programmer: M. Breitenfeld + * December 31, 2018 + * + *------------------------------------------------------------------------- + */ +void +H5CX_set_mpio_rank0_bcast(hbool_t rank0_bcast) +{ + H5CX_node_t **head = H5CX_get_my_context(); /* Get the pointer to the head of the API context, for this thread */ + + FUNC_ENTER_NOAPI_NOINIT_NOERR + + /* Sanity checks */ + HDassert(head && *head); + + (*head)->ctx.rank0_bcast = rank0_bcast; + + FUNC_LEAVE_NOAPI_VOID +} /* end H5CX_set_mpio_rank0_bcast() */ #endif /* H5_HAVE_PARALLEL */ @@ -2596,6 +2653,40 @@ H5CX_test_set_mpio_coll_chunk_multi_ratio_ind(int mpio_coll_chunk_multi_ratio_in done: FUNC_LEAVE_NOAPI(ret_value) } /* end H5CX_test_set_mpio_coll_chunk_multi_ratio_ind() */ + + +/*------------------------------------------------------------------------- + * Function: H5CX_test_set_mpio_coll_rank0_bcast + * + * Purpose: Sets the instrumented "read-with-rank0-bcast" flag for the current API call context. + * + * Note: Only sets value if property set in DXPL + * + * Return: Non-negative on success / Negative on failure + * + * Programmer: Quincey Koziol + * January 2, 2019 + * + *------------------------------------------------------------------------- + */ +herr_t +H5CX_test_set_mpio_coll_rank0_bcast(hbool_t mpio_coll_rank0_bcast) +{ + H5CX_node_t **head = H5CX_get_my_context(); /* Get the pointer to the head of the API context, for this thread */ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI_NOINIT + + /* Sanity checks */ + HDassert(head && *head); + HDassert(!((*head)->ctx.dxpl_id == H5P_DEFAULT || + (*head)->ctx.dxpl_id == H5P_DATASET_XFER_DEFAULT)); + + H5CX_TEST_SET_PROP(H5D_XFER_COLL_RANK0_BCAST_NAME, mpio_coll_rank0_bcast) + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5CX_test_set_mpio_coll_rank0_bcast() */ #endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ #endif /* H5_HAVE_PARALLEL */ @@ -2640,6 +2731,7 @@ H5CX__pop_common(void) H5CX_SET_PROP(H5D_XFER_COLL_CHUNK_LINK_NUM_FALSE_NAME, mpio_coll_chunk_link_num_false) H5CX_SET_PROP(H5D_XFER_COLL_CHUNK_MULTI_RATIO_COLL_NAME, mpio_coll_chunk_multi_ratio_coll) H5CX_SET_PROP(H5D_XFER_COLL_CHUNK_MULTI_RATIO_IND_NAME, mpio_coll_chunk_multi_ratio_ind) + H5CX_SET_PROP(H5D_XFER_COLL_RANK0_BCAST_NAME, mpio_coll_rank0_bcast) #endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ #endif /* H5_HAVE_PARALLEL */ diff --git a/src/H5CXprivate.h b/src/H5CXprivate.h index 46289c4..46d25d0 100644 --- a/src/H5CXprivate.h +++ b/src/H5CXprivate.h @@ -77,6 +77,7 @@ H5_DLL H5AC_ring_t H5CX_get_ring(void); H5_DLL hbool_t H5CX_get_coll_metadata_read(void); H5_DLL herr_t H5CX_get_mpi_coll_datatypes(MPI_Datatype *btype, MPI_Datatype *ftype); H5_DLL hbool_t H5CX_get_mpi_file_flushing(void); +H5_DLL hbool_t H5CX_get_mpio_rank0_bcast(void); #endif /* H5_HAVE_PARALLEL */ /* "Getter" routines for DXPL properties cached in API context */ @@ -112,6 +113,7 @@ H5_DLL void H5CX_set_coll_metadata_read(hbool_t cmdr); H5_DLL herr_t H5CX_set_mpi_coll_datatypes(MPI_Datatype btype, MPI_Datatype ftype); H5_DLL herr_t H5CX_set_mpio_coll_opt(H5FD_mpio_collective_opt_t mpio_coll_opt); H5_DLL void H5CX_set_mpi_file_flushing(hbool_t flushing); +H5_DLL void H5CX_set_mpio_rank0_bcast(hbool_t rank0_bcast); #endif /* H5_HAVE_PARALLEL */ /* "Setter" routines for DXPL properties cached in API context */ @@ -137,6 +139,7 @@ H5_DLL herr_t H5CX_test_set_mpio_coll_chunk_link_num_true(int mpio_coll_chunk_li H5_DLL herr_t H5CX_test_set_mpio_coll_chunk_link_num_false(int mpio_coll_chunk_link_num_false); H5_DLL herr_t H5CX_test_set_mpio_coll_chunk_multi_ratio_coll(int mpio_coll_chunk_multi_ratio_coll); H5_DLL herr_t H5CX_test_set_mpio_coll_chunk_multi_ratio_ind(int mpio_coll_chunk_multi_ratio_ind); +H5_DLL herr_t H5CX_test_set_mpio_coll_rank0_bcast(hbool_t rank0_bcast); #endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ #endif /* H5_HAVE_PARALLEL */ diff --git a/src/H5Dio.c b/src/H5Dio.c index fe85d23..6062dff 100644 --- a/src/H5Dio.c +++ b/src/H5Dio.c @@ -657,22 +657,12 @@ H5D__write(H5D_t *dataset, hid_t mem_type_id, const H5S_t *mem_space, /* Various MPI based checks */ #ifdef H5_HAVE_PARALLEL - if H5F_HAS_FEATURE(dataset->oloc.file, H5FD_FEAT_HAS_MPI) { - /* If MPI based VFD is used, no VL datatype support yet. */ + if(H5F_HAS_FEATURE(dataset->oloc.file, H5FD_FEAT_HAS_MPI)) { + /* If MPI based VFD is used, no VL or region reference datatype support yet. */ /* This is because they use the global heap in the file and we don't */ /* support parallel access of that yet */ - if(H5T_detect_class(type_info.mem_type, H5T_VLEN, FALSE) > 0) - HGOTO_ERROR(H5E_DATASET, H5E_UNSUPPORTED, FAIL, "Parallel IO does not support writing VL datatypes yet") - - /* If MPI based VFD is used, no VL datatype support yet. */ - /* This is because they use the global heap in the file and we don't */ - /* support parallel access of that yet */ - /* We should really use H5T_detect_class() here, but it will be difficult - * to detect the type of the reference if it is nested... -QAK - */ - if(H5T_get_class(type_info.mem_type, TRUE) == H5T_REFERENCE && - H5T_get_ref_type(type_info.mem_type) == H5R_DATASET_REGION) - HGOTO_ERROR(H5E_DATASET, H5E_UNSUPPORTED, FAIL, "Parallel IO does not support writing region reference datatypes yet") + if(H5T_is_vl_storage(type_info.mem_type) > 0) + HGOTO_ERROR(H5E_DATASET, H5E_UNSUPPORTED, FAIL, "Parallel IO does not support writing VL or region reference datatypes yet") } /* end if */ else { H5FD_mpio_xfer_t io_xfer_mode; /* MPI I/O transfer mode */ diff --git a/src/H5Dmpio.c b/src/H5Dmpio.c index 2c06800..f5da33d 100644 --- a/src/H5Dmpio.c +++ b/src/H5Dmpio.c @@ -37,7 +37,6 @@ #include "H5Eprivate.h" /* Error handling */ #include "H5Fprivate.h" /* File access */ #include "H5FDprivate.h" /* File drivers */ -#include "H5FDmpi.h" /* MPI-based file drivers */ #include "H5Iprivate.h" /* IDs */ #include "H5MMprivate.h" /* Memory management */ #include "H5Oprivate.h" /* Object headers */ @@ -89,10 +88,20 @@ /******************/ /* Combine chunk address and chunk info into a struct for better performance. */ typedef struct H5D_chunk_addr_info_t { - haddr_t chunk_addr; - H5D_chunk_info_t chunk_info; + haddr_t chunk_addr; + H5D_chunk_info_t chunk_info; } H5D_chunk_addr_info_t; +/* Rank 0 Bcast values */ +typedef enum H5D_mpio_no_rank0_bcast_cause_t { + H5D_MPIO_RANK0_BCAST = 0x00, + H5D_MPIO_RANK0_NOT_H5S_ALL = 0x01, + H5D_MPIO_RANK0_NOT_CONTIGUOUS = 0x02, + H5D_MPIO_RANK0_NOT_FIXED_SIZE = 0x04, + H5D_MPIO_RANK0_GREATER_THAN_2GB = 0x08 +} H5D_mpio_no_rank0_bcast_cause_t; + + /* * Information about a single chunk when performing collective filtered I/O. All * of the fields of one of these structs are initialized at the start of collective @@ -281,8 +290,10 @@ H5D__mpio_opt_possible(const H5D_io_info_t *io_info, const H5S_t *file_space, const H5S_t *mem_space, const H5D_type_info_t *type_info) { H5FD_mpio_xfer_t io_xfer_mode; /* MPI I/O transfer mode */ - unsigned local_cause = 0; /* Local reason(s) for breaking collective mode */ - unsigned global_cause = 0; /* Global reason(s) for breaking collective mode */ + unsigned local_cause[2] = {0,0}; /* [0] Local reason(s) for breaking collective mode */ + /* [1] Flag if dataset is both: H5S_ALL and small */ + unsigned global_cause[2] = {0,0}; /* Global reason(s) for breaking collective mode */ + htri_t is_vl_storage; /* Whether the dataset's datatype is stored in a variable-length form */ htri_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_PACKAGE @@ -296,36 +307,37 @@ H5D__mpio_opt_possible(const H5D_io_info_t *io_info, const H5S_t *file_space, /* For independent I/O, get out quickly and don't try to form consensus */ if(H5CX_get_io_xfer_mode(&io_xfer_mode) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get MPI-I/O transfer mode") + /* Set error flag, but keep going */ + local_cause[0] |= H5D_MPIO_ERROR_WHILE_CHECKING_COLLECTIVE_POSSIBLE; if(io_xfer_mode == H5FD_MPIO_INDEPENDENT) - local_cause |= H5D_MPIO_SET_INDEPENDENT; + local_cause[0] |= H5D_MPIO_SET_INDEPENDENT; /* Optimized MPI types flag must be set */ /* (based on 'HDF5_MPI_OPT_TYPES' environment variable) */ if(!H5FD_mpi_opt_types_g) - local_cause |= H5D_MPIO_MPI_OPT_TYPES_ENV_VAR_DISABLED; + local_cause[0] |= H5D_MPIO_MPI_OPT_TYPES_ENV_VAR_DISABLED; /* Don't allow collective operations if datatype conversions need to happen */ if(!type_info->is_conv_noop) - local_cause |= H5D_MPIO_DATATYPE_CONVERSION; + local_cause[0] |= H5D_MPIO_DATATYPE_CONVERSION; /* Don't allow collective operations if data transform operations should occur */ if(!type_info->is_xform_noop) - local_cause |= H5D_MPIO_DATA_TRANSFORMS; + local_cause[0] |= H5D_MPIO_DATA_TRANSFORMS; /* Check whether these are both simple or scalar dataspaces */ if(!((H5S_SIMPLE == H5S_GET_EXTENT_TYPE(mem_space) || H5S_SCALAR == H5S_GET_EXTENT_TYPE(mem_space)) && (H5S_SIMPLE == H5S_GET_EXTENT_TYPE(file_space) || H5S_SCALAR == H5S_GET_EXTENT_TYPE(file_space)))) - local_cause |= H5D_MPIO_NOT_SIMPLE_OR_SCALAR_DATASPACES; + local_cause[0] |= H5D_MPIO_NOT_SIMPLE_OR_SCALAR_DATASPACES; /* Dataset storage must be contiguous or chunked */ if(!(io_info->dset->shared->layout.type == H5D_CONTIGUOUS || io_info->dset->shared->layout.type == H5D_CHUNKED)) - local_cause |= H5D_MPIO_NOT_CONTIGUOUS_OR_CHUNKED_DATASET; + local_cause[0] |= H5D_MPIO_NOT_CONTIGUOUS_OR_CHUNKED_DATASET; /* check if external-file storage is used */ if(io_info->dset->shared->dcpl_cache.efl.nused > 0) - local_cause |= H5D_MPIO_NOT_CONTIGUOUS_OR_CHUNKED_DATASET; + local_cause[0] |= H5D_MPIO_NOT_CONTIGUOUS_OR_CHUNKED_DATASET; /* The handling of memory space is different for chunking and contiguous * storage. For contiguous storage, mem_space and file_space won't change @@ -340,31 +352,84 @@ H5D__mpio_opt_possible(const H5D_io_info_t *io_info, const H5S_t *file_space, * is less than 3. The functions needed (MPI_Mprobe and MPI_Imrecv) will * not be available. */ - if (io_info->op_type == H5D_IO_OP_WRITE && - io_info->dset->shared->layout.type == H5D_CHUNKED && - io_info->dset->shared->dcpl_cache.pline.nused > 0) - local_cause |= H5D_MPIO_PARALLEL_FILTERED_WRITES_DISABLED; + if(io_info->op_type == H5D_IO_OP_WRITE && + io_info->dset->shared->layout.type == H5D_CHUNKED && + io_info->dset->shared->dcpl_cache.pline.nused > 0) + local_cause[0] |= H5D_MPIO_PARALLEL_FILTERED_WRITES_DISABLED; #endif + /* Check if we are able to do a MPI_Bcast of the data from one rank + * instead of having all the processes involved in the collective I/O call. + */ + + /* Check to see if the process is reading the entire dataset */ + if(H5S_GET_SELECT_TYPE(file_space) != H5S_SEL_ALL) + local_cause[1] |= H5D_MPIO_RANK0_NOT_H5S_ALL; + /* Only perform this optimization for contigous datasets, currently */ + else if(H5D_CONTIGUOUS != io_info->dset->shared->layout.type) + /* Flag to do a MPI_Bcast of the data from one proc instead of + * having all the processes involved in the collective I/O. + */ + local_cause[1] |= H5D_MPIO_RANK0_NOT_CONTIGUOUS; + else if((is_vl_storage = H5T_is_vl_storage(type_info->dset_type)) < 0) + local_cause[0] |= H5D_MPIO_ERROR_WHILE_CHECKING_COLLECTIVE_POSSIBLE; + else if(is_vl_storage) + local_cause[1] |= H5D_MPIO_RANK0_NOT_FIXED_SIZE; + else { + size_t type_size; /* Size of dataset's datatype */ + + /* Retrieve the size of the dataset's datatype */ + if(0 == (type_size = H5T_GET_SIZE(type_info->dset_type))) + local_cause[0] |= H5D_MPIO_ERROR_WHILE_CHECKING_COLLECTIVE_POSSIBLE; + else { + hssize_t snelmts; /* [Signed] # of elements in dataset's dataspace */ + + /* Retrieve the size of the dataset's datatype */ + if((snelmts = H5S_GET_EXTENT_NPOINTS(file_space)) < 0) + local_cause[0] |= H5D_MPIO_ERROR_WHILE_CHECKING_COLLECTIVE_POSSIBLE; + else { + hsize_t dset_size; + + /* Determine dataset size */ + dset_size = ((hsize_t)snelmts) * type_size; + + /* If the size of the dataset is less than 2GB then do an MPI_Bcast + * of the data from one process instead of having all the processes + * involved in the collective I/O. + */ + if(dset_size > ((hsize_t)(2.0F * H5_GB) - 1)) + local_cause[1] |= H5D_MPIO_RANK0_GREATER_THAN_2GB; + } /* end else */ + } /* end else */ + } /* end else */ + /* Check for independent I/O */ - if(local_cause & H5D_MPIO_SET_INDEPENDENT) - global_cause = local_cause; + if(local_cause[0] & H5D_MPIO_SET_INDEPENDENT) + global_cause[0] = local_cause[0]; else { int mpi_code; /* MPI error code */ /* Form consensus opinion among all processes about whether to perform * collective I/O */ - if(MPI_SUCCESS != (mpi_code = MPI_Allreduce(&local_cause, &global_cause, 1, MPI_UNSIGNED, MPI_BOR, io_info->comm))) + if(MPI_SUCCESS != (mpi_code = MPI_Allreduce(&local_cause, &global_cause, 2, MPI_UNSIGNED, MPI_BOR, io_info->comm))) HMPI_GOTO_ERROR(FAIL, "MPI_Allreduce failed", mpi_code) } /* end else */ /* Set the local & global values of no-collective-cause in the API context */ - H5CX_set_mpio_local_no_coll_cause(local_cause); - H5CX_set_mpio_global_no_coll_cause(global_cause); + H5CX_set_mpio_local_no_coll_cause(local_cause[0]); + H5CX_set_mpio_global_no_coll_cause(global_cause[0]); + + /* Set read-with-rank0-and-bcast flag if possible */ + if(global_cause[0] == 0 && global_cause[1] == 0) { + H5CX_set_mpio_rank0_bcast(TRUE); +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + H5CX_test_set_mpio_coll_rank0_bcast(TRUE); +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + } /* end if */ /* Set the return value, based on the global cause */ - ret_value = global_cause > 0 ? FALSE : TRUE; + ret_value = global_cause[0] > 0 ? FALSE : TRUE; done: FUNC_LEAVE_NOAPI(ret_value) @@ -3069,8 +3134,8 @@ H5D__filtered_collective_chunk_entry_io(H5D_filtered_collective_io_info_t *chunk chunk_entry->chunk_states.new_chunk.length = chunk_entry->chunk_states.chunk_current.length; /* Currently, these chunk reads are done independently and will likely - * cause issues with collective metadata reads enabled. In the future, - * this should be refactored to use collective chunk reads - JTH */ + * cause issues with collective metadata reads enabled. In the future, + * this should be refactored to use collective chunk reads - JTH */ /* Get the original state of parallel I/O transfer mode */ if(H5CX_get_io_xfer_mode(&xfer_mode) < 0) diff --git a/src/H5Dprivate.h b/src/H5Dprivate.h index aaa3db2..6fb7889 100644 --- a/src/H5Dprivate.h +++ b/src/H5Dprivate.h @@ -95,7 +95,13 @@ /* Definitions for all collective chunk instrumentation properties */ #define H5D_XFER_COLL_CHUNK_SIZE sizeof(unsigned) #define H5D_XFER_COLL_CHUNK_DEF 1 -#define H5D_XFER_COLL_CHUNK_FIX 0 + +/* General collective I/O instrumentation properties */ +#define H5D_XFER_COLL_RANK0_BCAST_NAME "coll_rank0_bcast" + +/* Definitions for general collective I/O instrumentation properties */ +#define H5D_XFER_COLL_RANK0_BCAST_SIZE sizeof(hbool_t) +#define H5D_XFER_COLL_RANK0_BCAST_DEF FALSE #endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ /* Default temporary buffer size */ diff --git a/src/H5FDmpio.c b/src/H5FDmpio.c index d160858..3ab90aa 100644 --- a/src/H5FDmpio.c +++ b/src/H5FDmpio.c @@ -1354,6 +1354,7 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, int n; #endif hbool_t use_view_this_time = FALSE; + hbool_t rank0_bcast = FALSE; /* If read-with-rank0-and-bcast flag was used */ herr_t ret_value = SUCCEED; FUNC_ENTER_STATIC @@ -1437,8 +1438,25 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, if(H5FD_mpio_Debug[(int)'r']) HDfprintf(stdout, "%s: doing MPI collective IO\n", FUNC); #endif - if(MPI_SUCCESS != (mpi_code = MPI_File_read_at_all(file->f, mpi_off, buf, size_i, buf_type, &mpi_stat))) - HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at_all failed", mpi_code) + /* Check whether we should read from rank 0 and broadcast to other ranks */ + if(H5CX_get_mpio_rank0_bcast()) { +#ifdef H5FDmpio_DEBUG + if(H5FD_mpio_Debug[(int)'r']) + HDfprintf(stdout, "%s: doing read-rank0-and-MPI_Bcast\n", FUNC); +#endif + /* Indicate path we've taken */ + rank0_bcast = TRUE; + + /* Read on rank 0 Bcast to other ranks */ + if(file->mpi_rank == 0) + if(MPI_SUCCESS != (mpi_code = MPI_File_read_at(file->f, mpi_off, buf, size_i, buf_type, &mpi_stat))) + HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at failed", mpi_code) + if(MPI_SUCCESS != (mpi_code = MPI_Bcast(buf, size_i, buf_type, 0, file->comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_code) + } /* end if */ + else + if(MPI_SUCCESS != (mpi_code = MPI_File_read_at_all(file->f, mpi_off, buf, size_i, buf_type, &mpi_stat))) + HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at_all failed", mpi_code) } /* end if */ else { #ifdef H5FDmpio_DEBUG @@ -1460,13 +1478,26 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, if(MPI_SUCCESS != (mpi_code = MPI_File_read_at(file->f, mpi_off, buf, size_i, buf_type, &mpi_stat))) HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at failed", mpi_code) - /* How many bytes were actually read? */ + /* Only retrieve bytes read if this rank _actually_ participated in I/O */ + if(!rank0_bcast || (rank0_bcast && file->mpi_rank == 0) ) { + /* How many bytes were actually read? */ #if MPI_VERSION >= 3 - if(MPI_SUCCESS != (mpi_code = MPI_Get_elements_x(&mpi_stat, buf_type, &bytes_read))) + if(MPI_SUCCESS != (mpi_code = MPI_Get_elements_x(&mpi_stat, buf_type, &bytes_read))) #else - if(MPI_SUCCESS != (mpi_code = MPI_Get_elements(&mpi_stat, MPI_BYTE, &bytes_read))) + if(MPI_SUCCESS != (mpi_code = MPI_Get_elements(&mpi_stat, MPI_BYTE, &bytes_read))) #endif - HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements failed", mpi_code) + HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements failed", mpi_code) + } /* end if */ + + /* If the rank0-bcast feature was used, broadcast the # of bytes read to + * other ranks, which didn't perform any I/O. + */ + /* NOTE: This could be optimized further to be combined with the broadcast + * of the data. (QAK - 2019/1/2) + */ + if(rank0_bcast) + if(MPI_SUCCESS != MPI_Bcast(&bytes_read, 1, MPI_LONG_LONG, 0, file->comm)) + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", 0) /* Get the type's size */ #if MPI_VERSION >= 3 diff --git a/src/H5Pdxpl.c b/src/H5Pdxpl.c index bfc1d93..8338d84 100644 --- a/src/H5Pdxpl.c +++ b/src/H5Pdxpl.c @@ -2017,6 +2017,7 @@ done: FUNC_LEAVE_API(ret_value) } /* end H5Pget_mpio_actual_io_mode() */ + /*------------------------------------------------------------------------- * Function: H5Pget_mpio_no_collective_cause * @@ -2053,8 +2054,6 @@ H5Pget_mpio_no_collective_cause(hid_t plist_id, uint32_t *local_no_collective_ca done: FUNC_LEAVE_API(ret_value) } /* end H5Pget_mpio_no_collective_cause() */ - - #endif /* H5_HAVE_PARALLEL */ diff --git a/src/H5Ppublic.h b/src/H5Ppublic.h index 2f094ea..078fe74 100644 --- a/src/H5Ppublic.h +++ b/src/H5Ppublic.h @@ -167,7 +167,8 @@ typedef enum H5D_mpio_no_collective_cause_t { H5D_MPIO_NOT_SIMPLE_OR_SCALAR_DATASPACES = 0x10, H5D_MPIO_NOT_CONTIGUOUS_OR_CHUNKED_DATASET = 0x20, H5D_MPIO_PARALLEL_FILTERED_WRITES_DISABLED = 0x40, - H5D_MPIO_NO_COLLECTIVE_MAX_CAUSE = 0x80 + H5D_MPIO_ERROR_WHILE_CHECKING_COLLECTIVE_POSSIBLE = 0x80, + H5D_MPIO_NO_COLLECTIVE_MAX_CAUSE = 0x100 } H5D_mpio_no_collective_cause_t; /********************/ diff --git a/src/H5T.c b/src/H5T.c index 01ace87..9544488 100644 --- a/src/H5T.c +++ b/src/H5T.c @@ -295,6 +295,7 @@ static htri_t H5T__compiler_conv(H5T_t *src, H5T_t *dst); static herr_t H5T__set_size(H5T_t *dt, size_t size); static herr_t H5T__close_cb(H5T_t *dt); static H5T_path_t *H5T__path_find_real(const H5T_t *src, const H5T_t *dst, const char *name, H5T_conv_func_t *conv); +static hbool_t H5T__detect_reg_ref(const H5T_t *dt); /*****************************/ @@ -5506,6 +5507,111 @@ done: /*------------------------------------------------------------------------- + * Function: H5T_detect_reg_ref + * + * Purpose: Check whether a datatype contains (or is) a region reference + * datatype. + * + * Return: TRUE (1) or FALSE (0) on success + * (Can't fail) + * + * Programmer: Quincey Koziol + * Saturday, January 5, 2019 + * + *------------------------------------------------------------------------- + */ +static hbool_t +H5T__detect_reg_ref(const H5T_t *dt) +{ + unsigned u; /* Local index variable */ + hbool_t ret_value = FALSE; /* Return value */ + + FUNC_ENTER_STATIC_NOERR + + /* Sanity checks */ + HDassert(dt); + + /* Check if this datatype is a region reference */ + if(H5T_REFERENCE == dt->shared->type && H5R_DATASET_REGION == dt->shared->u.atomic.u.r.rtype) + HGOTO_DONE(TRUE); + + /* Check for types that might have the correct type as a component */ + switch(dt->shared->type) { + case H5T_COMPOUND: + /* Iterate over all the compound datatype's fields */ + for(u = 0; u < dt->shared->u.compnd.nmembs; u++) + /* Recurse on field's datatype */ + if(H5T__detect_reg_ref(dt->shared->u.compnd.memb[u].type)) + HGOTO_DONE(TRUE); + break; + + case H5T_ARRAY: + case H5T_VLEN: + case H5T_ENUM: + HGOTO_DONE(H5T__detect_reg_ref(dt->shared->parent)); + break; + + case H5T_NO_CLASS: + case H5T_INTEGER: + case H5T_FLOAT: + case H5T_TIME: + case H5T_STRING: + case H5T_BITFIELD: + case H5T_OPAQUE: + case H5T_REFERENCE: + case H5T_NCLASSES: + default: + break; + } /* end if */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5T__detect_reg_ref() */ + + +/*------------------------------------------------------------------------- + * Function: H5T_is_vl_storage + * + * Purpose: Check if a datatype will be stored in a variable-length form. + * + * Notes: Currently, only variable-length string & sequences and region + * references are stored in a variable-length form. + * + * Return: + * One of two values on success: + * TRUE - If the datatype will be stored in a variable-length form + * FALSE - If the datatype will NOT be stored in a variable-length form + * <0 is returned on failure + * + * Programmer: Quincey Koziol + * Saturday, January 5, 2019 + * + *------------------------------------------------------------------------- + */ +htri_t +H5T_is_vl_storage(const H5T_t *dt) +{ + htri_t ret_value = FALSE; + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity check */ + HDassert(dt); + + /* VL and region reference datatypes are stored in variable-length form */ + if(H5T_detect_class(dt, H5T_VLEN, FALSE)) + ret_value = TRUE; + else if(H5T_detect_class(dt, H5T_REFERENCE, FALSE)) + ret_value = H5T__detect_reg_ref(dt); + else + ret_value = FALSE; + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5T_is_vl_storage() */ + + +/*------------------------------------------------------------------------- * Function: H5T_upgrade_version_cb * * Purpose: H5T__visit callback to Upgrade the version of a datatype diff --git a/src/H5Tprivate.h b/src/H5Tprivate.h index 6b6446f..3dcbb2c 100644 --- a/src/H5Tprivate.h +++ b/src/H5Tprivate.h @@ -145,6 +145,7 @@ H5_DLL H5T_t *H5T_get_actual_type(H5T_t *dt); H5_DLL herr_t H5T_save_refresh_state(hid_t tid, struct H5O_shared_t *cached_H5O_shared); H5_DLL herr_t H5T_restore_refresh_state(hid_t tid, struct H5O_shared_t *cached_H5O_shared); H5_DLL hbool_t H5T_already_vol_managed(const H5T_t *dt); +H5_DLL htri_t H5T_is_vl_storage(const H5T_t *dt); /* Reference specific functions */ H5_DLL H5R_type_t H5T_get_ref_type(const H5T_t *dt); diff --git a/testpar/t_bigio.c b/testpar/t_bigio.c index fdd3488..1d882b8 100644 --- a/testpar/t_bigio.c +++ b/testpar/t_bigio.c @@ -671,7 +671,7 @@ dataset_big_write(void) /* create a memory dataspace independently */ mem_dataspace = H5Screate_simple (RANK, dims, NULL); VRFY((mem_dataspace >= 0), ""); - if(!mpi_rank == 0) { + if(mpi_rank != 0) { ret = H5Sselect_none(mem_dataspace); VRFY((ret >= 0), "H5Sset_none succeeded"); } @@ -980,7 +980,7 @@ dataset_big_read(void) /* create a memory dataspace independently */ mem_dataspace = H5Screate_simple (RANK, dims, NULL); VRFY((mem_dataspace >= 0), ""); - if(!mpi_rank == 0) { + if(mpi_rank != 0) { ret = H5Sselect_none(mem_dataspace); VRFY((ret >= 0), "H5Sset_none succeeded"); } diff --git a/testpar/t_dset.c b/testpar/t_dset.c index 281d027..35501d8 100644 --- a/testpar/t_dset.c +++ b/testpar/t_dset.c @@ -2649,7 +2649,7 @@ compress_readAll(void) /* Try reading the data */ ret = H5Dread(dataset, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, xfer_plist, data_read); - VRFY((ret >= 0), "H5Pset_dxpl_mpio succeeded"); + VRFY((ret >= 0), "H5Dread succeeded"); /* Verify data read */ for(u=0; u= 0),"set independent IO collectively succeeded"); } - dataset = H5Dopen2(iof, dname, H5P_DEFAULT); VRFY((dataset >= 0), "H5Dopen2 succeeded"); +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + hbool_t prop_value; + prop_value = H5D_XFER_COLL_RANK0_BCAST_DEF; + ret = H5Pinsert2(dxpl, H5D_XFER_COLL_RANK0_BCAST_NAME, H5D_XFER_COLL_RANK0_BCAST_SIZE, &prop_value, + NULL, NULL, NULL, NULL, NULL, NULL); + VRFY((ret >= 0), "H5Pinsert2() succeeded"); +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + ret = H5Dread(dataset, H5T_NATIVE_DOUBLE, H5S_ALL, H5S_ALL, dxpl, inme); VRFY((ret >= 0), "H5Dread succeeded"); +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + prop_value = FALSE; + ret = H5Pget(dxpl, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value); + VRFY((ret >= 0), "H5Pget succeeded"); + VRFY((prop_value == FALSE && dxfer_coll_type == DXFER_COLLECTIVE_IO),"rank 0 Bcast optimization was performed for a compact dataset"); +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + /* Verify data value */ for(i = 0; i < size; i++) for(j = 0; j < size; j++) @@ -603,8 +618,8 @@ void dataset_fillvalue(void) hsize_t req_count[4] = {1, 6, 7, 8}; hsize_t dset_size; /* Dataset size */ int *rdata, *wdata; /* Buffers for data to read and write */ - int *twdata, *trdata; /* Temporary pointer into buffer */ - int acc, i, j, k, l; /* Local index variables */ + int *twdata, *trdata; /* Temporary pointer into buffer */ + int acc, i, j, k, l, ii; /* Local index variables */ herr_t ret; /* Generic return value */ const char *filename; @@ -645,27 +660,60 @@ void dataset_fillvalue(void) /* * Read dataset before any data is written. */ - /* set entire read buffer with the constant 2 */ - HDmemset(rdata,2,(size_t)(dset_size*sizeof(int))); - /* Independently read the entire dataset back */ - ret = H5Dread(dataset, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, H5P_DEFAULT, rdata); - VRFY((ret >= 0), "H5Dread succeeded"); - /* Verify all data read are the fill value 0 */ - trdata = rdata; - err_num = 0; - for(i = 0; i < (int)dset_dims[0]; i++) + /* Create DXPL for I/O */ + dxpl = H5Pcreate(H5P_DATASET_XFER); + VRFY((dxpl >= 0), "H5Pcreate succeeded"); + +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + hbool_t prop_value; + prop_value = H5D_XFER_COLL_RANK0_BCAST_DEF; + ret = H5Pinsert2(dxpl, H5D_XFER_COLL_RANK0_BCAST_NAME, H5D_XFER_COLL_RANK0_BCAST_SIZE, &prop_value, + NULL, NULL, NULL, NULL, NULL, NULL); + VRFY((ret >= 0),"testing property list inserted succeeded"); +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + + for(ii = 0; ii < 2; ii++) { + + if(ii == 0) + ret = H5Pset_dxpl_mpio(dxpl, H5FD_MPIO_INDEPENDENT); + else + ret = H5Pset_dxpl_mpio(dxpl, H5FD_MPIO_COLLECTIVE); + VRFY((ret >= 0), "H5Pset_dxpl_mpio succeeded"); + + /* set entire read buffer with the constant 2 */ + HDmemset(rdata,2,(size_t)(dset_size*sizeof(int))); + + /* Read the entire dataset back */ + ret = H5Dread(dataset, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, dxpl, rdata); + VRFY((ret >= 0), "H5Dread succeeded"); + +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + prop_value = FALSE; + ret = H5Pget(dxpl, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value); + VRFY((ret >= 0), "testing property list get succeeded"); + if(ii == 0) + VRFY((prop_value == FALSE), "correctly handled rank 0 Bcast"); + else + VRFY((prop_value == TRUE), "correctly handled rank 0 Bcast"); +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + + /* Verify all data read are the fill value 0 */ + trdata = rdata; + err_num = 0; + for(i = 0; i < (int)dset_dims[0]; i++) for(j = 0; j < (int)dset_dims[1]; j++) - for(k = 0; k < (int)dset_dims[2]; k++) - for(l = 0; l < (int)dset_dims[3]; l++, twdata++, trdata++) - if(*trdata != 0) - if(err_num++ < MAX_ERR_REPORT || VERBOSE_MED) - printf("Dataset Verify failed at [%d][%d][%d][%d]: expect 0, got %d\n", i, j, k, l, *trdata); - if(err_num > MAX_ERR_REPORT && !VERBOSE_MED) + for(k = 0; k < (int)dset_dims[2]; k++) + for(l = 0; l < (int)dset_dims[3]; l++, twdata++, trdata++) + if(*trdata != 0) + if(err_num++ < MAX_ERR_REPORT || VERBOSE_MED) + printf("Dataset Verify failed at [%d][%d][%d][%d]: expect 0, got %d\n", i, j, k, l, *trdata); + if(err_num > MAX_ERR_REPORT && !VERBOSE_MED) printf("[more errors ...]\n"); - if(err_num){ + if(err_num) { printf("%d errors found in check_value\n", err_num); - nerrors++; + nerrors++; + } } /* Barrier to ensure all processes have completed the above test. */ @@ -681,10 +729,6 @@ void dataset_fillvalue(void) ret = H5Sselect_hyperslab(memspace, H5S_SELECT_SET, req_start, NULL, req_count, NULL); VRFY((ret >= 0), "H5Sselect_hyperslab succeeded on memory dataspace"); - /* Create DXPL for collective I/O */ - dxpl = H5Pcreate(H5P_DATASET_XFER); - VRFY((dxpl >= 0), "H5Pcreate succeeded"); - ret = H5Pset_dxpl_mpio(dxpl, H5FD_MPIO_COLLECTIVE); VRFY((ret >= 0), "H5Pset_dxpl_mpio succeeded"); if(dxfer_coll_type == DXFER_INDEPENDENT_IO) { @@ -711,37 +755,64 @@ void dataset_fillvalue(void) /* * Read dataset after partial write. */ - /* set entire read buffer with the constant 2 */ - HDmemset(rdata,2,(size_t)(dset_size*sizeof(int))); - /* Independently read the entire dataset back */ - ret = H5Dread(dataset, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, H5P_DEFAULT, rdata); - VRFY((ret >= 0), "H5Dread succeeded"); - /* Verify correct data read */ - twdata=wdata; - trdata=rdata; - err_num=0; - for(i=0; i<(int)dset_dims[0]; i++) +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + prop_value = H5D_XFER_COLL_RANK0_BCAST_DEF; + ret = H5Pset(dxpl, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value); + VRFY((ret >= 0), " H5Pset succeeded"); +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + + for(ii = 0; ii < 2; ii++) { + + if(ii == 0) + ret = H5Pset_dxpl_mpio(dxpl, H5FD_MPIO_INDEPENDENT); + else + ret = H5Pset_dxpl_mpio(dxpl, H5FD_MPIO_COLLECTIVE); + VRFY((ret >= 0), "H5Pset_dxpl_mpio succeeded"); + + /* set entire read buffer with the constant 2 */ + HDmemset(rdata,2,(size_t)(dset_size*sizeof(int))); + + /* Read the entire dataset back */ + ret = H5Dread(dataset, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, dxpl, rdata); + VRFY((ret >= 0), "H5Dread succeeded"); + +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + prop_value = FALSE; + ret = H5Pget(dxpl, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value); + VRFY((ret >= 0), "testing property list get succeeded"); + if(ii == 0) + VRFY((prop_value == FALSE), "correctly handled rank 0 Bcast"); + else + VRFY((prop_value == TRUE), "correctly handled rank 0 Bcast"); +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + + /* Verify correct data read */ + twdata=wdata; + trdata=rdata; + err_num=0; + for(i=0; i<(int)dset_dims[0]; i++) for(j=0; j<(int)dset_dims[1]; j++) - for(k=0; k<(int)dset_dims[2]; k++) - for(l=0; l<(int)dset_dims[3]; l++, twdata++, trdata++) - if(i MAX_ERR_REPORT && !VERBOSE_MED) + for(k=0; k<(int)dset_dims[2]; k++) + for(l=0; l<(int)dset_dims[3]; l++, twdata++, trdata++) + if(i MAX_ERR_REPORT && !VERBOSE_MED) printf("[more errors ...]\n"); - if(err_num){ + if(err_num){ printf("%d errors found in check_value\n", err_num); - nerrors++; + nerrors++; + } } - + /* Close all file objects */ ret = H5Dclose(dataset); VRFY((ret >= 0), "H5Dclose succeeded"); @@ -856,7 +927,7 @@ void collective_group_write(void) if(!((m+1) % 10)) { printf("created %d groups\n", m+1); MPI_Barrier(MPI_COMM_WORLD); - } + } #endif /* BARRIER_CHECKS */ } diff --git a/testpar/t_pread.c b/testpar/t_pread.c index 0905d44..74feeb6 100644 --- a/testpar/t_pread.c +++ b/testpar/t_pread.c @@ -17,6 +17,7 @@ */ #include "testpar.h" +#include "H5Dprivate.h" /* The collection of files is included below to aid * an external "cleanup" process if required. @@ -34,6 +35,8 @@ const char *FILENAMES[NFILENAME + 1]={"reloc_t_pread_data_file", #define COUNT 1000 +#define LIMIT_NPROC 6 + hbool_t pass = true; static const char *random_hdf5_text = "Now is the time for all first-time-users of HDF5 to read their \ @@ -46,7 +49,7 @@ completely foolproof is to underestimate the ingenuity of complete\n\ fools.\n"; static int generate_test_file(MPI_Comm comm, int mpi_rank, int group); -static int test_parallel_read(MPI_Comm comm, int mpi_rank, int group); +static int test_parallel_read(MPI_Comm comm, int mpi_rank, int mpi_size, int group); static char *test_argv0 = NULL; @@ -108,6 +111,9 @@ generate_test_file( MPI_Comm comm, int mpi_rank, int group_id ) hid_t fapl_id = -1; hid_t dxpl_id = -1; hid_t dset_id = -1; + hid_t dset_id_ch = -1; + hid_t dcpl_id = H5P_DEFAULT; + hsize_t chunk[1]; float nextValue; float *data_slice = NULL; @@ -272,6 +278,55 @@ generate_test_file( MPI_Comm comm, int mpi_rank, int group_id ) } } + + /* create a chunked dataset */ + chunk[0] = COUNT/8; + + if ( pass ) { + if ( (dcpl_id = H5Pcreate (H5P_DATASET_CREATE)) < 0 ) { + pass = false; + failure_mssg = "H5Pcreate() failed.\n"; + } + } + + if ( pass ) { + if ( (H5Pset_chunk (dcpl_id, 1, chunk) ) < 0 ) { + pass = false; + failure_mssg = "H5Pset_chunk() failed.\n"; + } + } + + if ( pass ) { + + if ( (dset_id_ch = H5Dcreate2(file_id, "dataset0_chunked", H5T_NATIVE_FLOAT, + filespace, H5P_DEFAULT, dcpl_id, + H5P_DEFAULT)) < 0 ) { + pass = false; + failure_mssg = "H5Dcreate2() failed.\n"; + } + } + + if ( pass ) { + if ( (H5Dwrite(dset_id_ch, H5T_NATIVE_FLOAT, memspace, + filespace, dxpl_id, data_slice)) < 0 ) { + pass = false; + failure_mssg = "H5Dwrite() failed.\n"; + } + } + if ( pass || (dcpl_id != -1)) { + if ( H5Pclose(dcpl_id) < 0 ) { + pass = false; + failure_mssg = "H5Pclose(dcpl_id) failed.\n"; + } + } + + if ( pass || (dset_id_ch != -1)) { + if ( H5Dclose(dset_id_ch) < 0 ) { + pass = false; + failure_mssg = "H5Dclose(dset_id_ch) failed.\n"; + } + } + /* close file, etc. */ if ( pass || (dset_id != -1)) { if ( H5Dclose(dset_id) < 0 ) { @@ -413,7 +468,7 @@ generate_test_file( MPI_Comm comm, int mpi_rank, int group_id ) * Function: test_parallel_read * * Purpose: This actually tests the superblock optimization - * and covers the two primary cases we're interested in. + * and covers the three primary cases we're interested in. * 1). That HDF5 files can be opened in parallel by * the rank 0 process and that the superblock * offset is correctly broadcast to the other @@ -423,6 +478,10 @@ generate_test_file( MPI_Comm comm, int mpi_rank, int group_id ) * subgroups of MPI_COMM_WORLD and that each * subgroup operates as described in (1) to * collectively read the data. + * 3). Testing proc0-read-and-MPI_Bcast using + * sub-communicators, and reading into + * a memory space that is different from the + * file space, and chunked datasets. * * The global MPI rank is used for reading and * writing data for process specific data in the @@ -444,7 +503,7 @@ generate_test_file( MPI_Comm comm, int mpi_rank, int group_id ) *------------------------------------------------------------------------- */ static int -test_parallel_read(MPI_Comm comm, int mpi_rank, int group_id) +test_parallel_read(MPI_Comm comm, int mpi_rank, int mpi_size, int group_id) { const char *failure_mssg; const char *fcn_name = "test_parallel_read()"; @@ -457,8 +516,13 @@ test_parallel_read(MPI_Comm comm, int mpi_rank, int group_id) hid_t fapl_id = -1; hid_t file_id = -1; hid_t dset_id = -1; + hid_t dset_id_ch = -1; + hid_t dxpl_id = H5P_DEFAULT; hid_t memspace = -1; hid_t filespace = -1; + hid_t filetype = -1; + size_t filetype_size; + hssize_t dset_size; hsize_t i; hsize_t offset; hsize_t count = COUNT; @@ -552,6 +616,14 @@ test_parallel_read(MPI_Comm comm, int mpi_rank, int group_id) } } + /* open the chunked data set */ + if ( pass ) { + if ( (dset_id_ch = H5Dopen2(file_id, "dataset0_chunked", H5P_DEFAULT)) < 0 ) { + pass = FALSE; + failure_mssg = "H5Dopen2() failed\n"; + } + } + /* setup memspace */ if ( pass ) { dims[0] = count; @@ -606,14 +678,6 @@ test_parallel_read(MPI_Comm comm, int mpi_rank, int group_id) } } - /* close file, etc. */ - if ( pass || (dset_id != -1) ) { - if ( H5Dclose(dset_id) < 0 ) { - pass = false; - failure_mssg = "H5Dclose(dset_id) failed.\n"; - } - } - if ( pass || (memspace != -1) ) { if ( H5Sclose(memspace) < 0 ) { pass = false; @@ -628,6 +692,330 @@ test_parallel_read(MPI_Comm comm, int mpi_rank, int group_id) } } + /* free data_slice if it has been allocated */ + if ( data_slice != NULL ) { + HDfree(data_slice); + data_slice = NULL; + } + + /* + * Test reading proc0-read-and-bcast with sub-communicators + */ + + /* Don't test with more than LIMIT_NPROC processes to avoid memory issues */ + + if( group_size <= LIMIT_NPROC ) { +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + hbool_t prop_value; +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + + if ( (filespace = H5Dget_space(dset_id )) < 0 ) { + pass = FALSE; + failure_mssg = "H5Dget_space failed.\n"; + } + + if ( (dset_size = H5Sget_simple_extent_npoints(filespace)) < 0 ) { + pass = FALSE; + failure_mssg = "H5Sget_simple_extent_npoints failed.\n"; + } + + if ( (filetype = H5Dget_type(dset_id)) < 0 ) { + pass = FALSE; + failure_mssg = "H5Dget_type failed.\n"; + } + + if ( (filetype_size = H5Tget_size(filetype)) == 0 ) { + pass = FALSE; + failure_mssg = "H5Tget_size failed.\n"; + } + + if ( H5Tclose(filetype) < 0 ) { + pass = FALSE; + failure_mssg = "H5Tclose failed.\n"; + }; + + if ( (data_slice = (float *)HDmalloc((size_t)dset_size*filetype_size)) == NULL ) { + pass = FALSE; + failure_mssg = "malloc of data_slice failed.\n"; + } + + if ( pass ) { + if ( (dxpl_id = H5Pcreate(H5P_DATASET_XFER)) < 0 ) { + pass = FALSE; + failure_mssg = "H5Pcreate(H5P_DATASET_XFER) failed.\n"; + } + } + + if ( pass ) { + if ( (H5Pset_dxpl_mpio(dxpl_id, H5FD_MPIO_COLLECTIVE)) < 0 ) { + pass = FALSE; + failure_mssg = "H5Pset_dxpl_mpio() failed.\n"; + } + } + +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + if ( pass ) { + prop_value = H5D_XFER_COLL_RANK0_BCAST_DEF; + if(H5Pinsert2(dxpl_id, H5D_XFER_COLL_RANK0_BCAST_NAME, H5D_XFER_COLL_RANK0_BCAST_SIZE, &prop_value, + NULL, NULL, NULL, NULL, NULL, NULL) < 0) { + pass = FALSE; + failure_mssg = "H5Pinsert2() failed\n"; + } + } +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + + /* read H5S_ALL section */ + if ( pass ) { + if ( (H5Dread(dset_id, H5T_NATIVE_FLOAT, H5S_ALL, + H5S_ALL, dxpl_id, data_slice)) < 0 ) { + pass = FALSE; + failure_mssg = "H5Dread() failed\n"; + } + } + +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + if ( pass ) { + prop_value = FALSE; + if(H5Pget(dxpl_id, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value) < 0) { + pass = FALSE; + failure_mssg = "H5Pget() failed\n"; + } + if (pass) { + if(prop_value != TRUE) { + pass = FALSE; + failure_mssg = "rank 0 Bcast optimization was mistakenly not performed\n"; + } + } + } +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + + /* verify the data */ + if ( pass ) { + + if ( comm == MPI_COMM_WORLD ) /* test 1 */ + nextValue = 0; + else if ( group_id == 0 ) /* test 2 group 0 */ + nextValue = 0; + else /* test 2 group 1 */ + nextValue = (float)((hsize_t)( mpi_size / 2 )*count); + + i = 0; + while ( ( pass ) && ( i < (hsize_t)dset_size ) ) { + /* what we really want is data_slice[i] != nextValue -- + * the following is a circumlocution to shut up the + * the compiler. + */ + if ( ( data_slice[i] > nextValue ) || + ( data_slice[i] < nextValue ) ) { + pass = FALSE; + failure_mssg = "Unexpected dset contents.\n"; + } + nextValue += 1; + i++; + } + } + + /* read H5S_ALL section for the chunked dataset */ + +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + if ( pass ) { + prop_value = H5D_XFER_COLL_RANK0_BCAST_DEF; + if(H5Pset(dxpl_id, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value) < 0) { + pass = FALSE; + failure_mssg = "H5Pset() failed\n"; + } + } +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + + for ( i = 0; i < (hsize_t)dset_size; i++) { + data_slice[i] = 0; + } + if ( pass ) { + if ( (H5Dread(dset_id_ch, H5T_NATIVE_FLOAT, H5S_ALL, + H5S_ALL, dxpl_id, data_slice)) < 0 ) { + pass = FALSE; + failure_mssg = "H5Dread() failed\n"; + } + } + +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + if ( pass ) { + prop_value = FALSE; + if(H5Pget(dxpl_id, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value) < 0) { + pass = FALSE; + failure_mssg = "H5Pget() failed\n"; + } + if (pass) { + if(prop_value == TRUE) { + pass = FALSE; + failure_mssg = "rank 0 Bcast optimization was mistakenly performed for chunked dataset\n"; + } + } + } +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + + /* verify the data */ + if ( pass ) { + + if ( comm == MPI_COMM_WORLD ) /* test 1 */ + nextValue = 0; + else if ( group_id == 0 ) /* test 2 group 0 */ + nextValue = 0; + else /* test 2 group 1 */ + nextValue = (float)((hsize_t)( mpi_size / 2 )*count); + + i = 0; + while ( ( pass ) && ( i < (hsize_t)dset_size ) ) { + /* what we really want is data_slice[i] != nextValue -- + * the following is a circumlocution to shut up the + * the compiler. + */ + if ( ( data_slice[i] > nextValue ) || + ( data_slice[i] < nextValue ) ) { + pass = FALSE; + failure_mssg = "Unexpected chunked dset contents.\n"; + } + nextValue += 1; + i++; + } + } + + if ( pass || (filespace != -1) ) { + if ( H5Sclose(filespace) < 0 ) { + pass = false; + failure_mssg = "H5Sclose(filespace) failed.\n"; + } + } + + /* free data_slice if it has been allocated */ + if ( data_slice != NULL ) { + HDfree(data_slice); + data_slice = NULL; + } + + /* + * Read an H5S_ALL filespace into a hyperslab defined memory space + */ + + if ( (data_slice = (float *)HDmalloc((size_t)(dset_size*2)*filetype_size)) == NULL ) { + pass = FALSE; + failure_mssg = "malloc of data_slice failed.\n"; + } + + /* setup memspace */ + if ( pass ) { + dims[0] = (hsize_t)dset_size*2; + if ( (memspace = H5Screate_simple(1, dims, NULL)) < 0 ) { + pass = FALSE; + failure_mssg = "H5Screate_simple(1, dims, NULL) failed\n"; + } + } + if ( pass ) { + offset = (hsize_t)dset_size; + if ( (H5Sselect_hyperslab(memspace, H5S_SELECT_SET, + &offset, NULL, &offset, NULL)) < 0 ) { + pass = FALSE; + failure_mssg = "H5Sselect_hyperslab() failed\n"; + } + } + +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + if ( pass ) { + prop_value = H5D_XFER_COLL_RANK0_BCAST_DEF; + if(H5Pset(dxpl_id, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value) < 0) { + pass = FALSE; + failure_mssg = "H5Pset() failed\n"; + } + } +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + + /* read this processes section of the data */ + if ( pass ) { + if ( (H5Dread(dset_id, H5T_NATIVE_FLOAT, memspace, + H5S_ALL, dxpl_id, data_slice)) < 0 ) { + pass = FALSE; + failure_mssg = "H5Dread() failed\n"; + } + } + +#ifdef H5_HAVE_INSTRUMENTED_LIBRARY + if ( pass ) { + prop_value = FALSE; + if(H5Pget(dxpl_id, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value) < 0) { + pass = FALSE; + failure_mssg = "H5Pget() failed\n"; + } + if (pass) { + if(prop_value != TRUE) { + pass = FALSE; + failure_mssg = "rank 0 Bcast optimization was mistakenly not performed\n"; + } + } + } +#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */ + + /* verify the data */ + if ( pass ) { + + if ( comm == MPI_COMM_WORLD ) /* test 1 */ + nextValue = 0; + else if ( group_id == 0 ) /* test 2 group 0 */ + nextValue = 0; + else /* test 2 group 1 */ + nextValue = (float)((hsize_t)(mpi_size / 2)*count); + + i = (hsize_t)dset_size; + while ( ( pass ) && ( i < (hsize_t)dset_size ) ) { + /* what we really want is data_slice[i] != nextValue -- + * the following is a circumlocution to shut up the + * the compiler. + */ + if ( ( data_slice[i] > nextValue ) || + ( data_slice[i] < nextValue ) ) { + pass = FALSE; + failure_mssg = "Unexpected dset contents.\n"; + } + nextValue += 1; + i++; + } + } + + if ( pass || (memspace != -1) ) { + if ( H5Sclose(memspace) < 0 ) { + pass = false; + failure_mssg = "H5Sclose(memspace) failed.\n"; + } + } + + /* free data_slice if it has been allocated */ + if ( data_slice != NULL ) { + HDfree(data_slice); + data_slice = NULL; + } + + if ( pass || (dxpl_id != -1) ) { + if ( H5Pclose(dxpl_id) < 0 ) { + pass = false; + failure_mssg = "H5Pclose(dxpl_id) failed.\n"; + } + } + } + + /* close file, etc. */ + if ( pass || (dset_id != -1) ) { + if ( H5Dclose(dset_id) < 0 ) { + pass = false; + failure_mssg = "H5Dclose(dset_id) failed.\n"; + } + } + + if ( pass || (dset_id_ch != -1) ) { + if ( H5Dclose(dset_id_ch) < 0 ) { + pass = false; + failure_mssg = "H5Dclose(dset_id_ch) failed.\n"; + } + } + if ( pass || (file_id != -1) ) { if ( H5Fclose(file_id) < 0 ) { pass = false; @@ -668,17 +1056,9 @@ test_parallel_read(MPI_Comm comm, int mpi_rank, int group_id) HDfprintf(stdout, "%s: failure_mssg = \"%s\"\n", fcn_name, failure_mssg); } - HDremove(reloc_data_filename); } - /* free data_slice if it has been allocated */ - if ( data_slice != NULL ) { - HDfree(data_slice); - data_slice = NULL; - } - - return( ! pass ); } /* test_parallel_read() */ @@ -803,7 +1183,7 @@ main( int argc, char **argv) } /* Now read the generated test file (stil using MPI_COMM_WORLD) */ - nerrs += test_parallel_read( MPI_COMM_WORLD, mpi_rank, which_group); + nerrs += test_parallel_read( MPI_COMM_WORLD, mpi_rank, mpi_size, which_group); if ( nerrs > 0 ) { if ( mpi_rank == 0 ) { @@ -819,7 +1199,7 @@ main( int argc, char **argv) } /* run the 2nd set of tests */ - nerrs += test_parallel_read(group_comm, mpi_rank, which_group); + nerrs += test_parallel_read(group_comm, mpi_rank, mpi_size, which_group); if ( nerrs > 0 ) { if ( mpi_rank == 0 ) { -- cgit v0.12