HDF5 documents and links Introduction to HDF5 HDF5 User Guide	And in this document, the HDF5 Reference Manual H5 H5A H5D H5E H5F H5G H5I H5P H5R H5S H5T H5Z Tools Datatypes Collective Calls in Parallel
(PDF of complete manual formatted as print volume)

Collective Calling Requirements
in Parallel HDF5 Applications

Introduction

This is the initial sketch of a document addressing two major topics:

HDF5 functions that must be called collectively and when in a parallel computing environment
Properties that must be used in a coordinated manner in a parallel computing environment

The notes referenced in parantheses with many function names follow the lists of functions and properties.

Always collective

The following functions must always be called collectively.

    H5Aclose (2)
    H5Acreate (6) (9)
    H5Adelete
    H5Arename (A)
    H5Awrite (3)

    H5Dclose (2)
    H5Dcreate (6) (9)
    H5Dextend (5) (10)
    H5Dset_extent (5) (10) (A)

    H5Fclose (1)
    H5Fcreate (8) (9)
    H5Fflush
    H5Fmount
    H5Fopen (9)
    H5Freopen
    H5Funmount

    H5Gclose (2)
    H5Gcreate (8) (9)
    H5Glink
    H5Glink2 (A)
    H5Gmove
    H5Gmove2 (A)
    H5Gset_comment
    H5Gunlink

    H5Idec_ref (7) (A)
    H5Iinc_ref (7) (A)

    H5Rcreate

    H5Tclose (4)
    H5Tcommit (8) (9)

Collective, unless target object will not be modified

The following functions must normally be called collectively. If, however, the target object will not be modified, they may be called independently.

    H5Aopen_idx (9)
    H5Aopen_name (9)

    H5Dopen (9)

    H5Gopen (9)

    H5Rdereference

    H5Topen (9)

Properties

The following properties must be set to the same values when they are used in a parallel program. Dataset creation properties:

    H5Pset_alloc_time
    H5Pset_chunk
    H5Pset_deflate
    H5Pset_external
    H5Pset_fill_time
    H5Pset_fill_value
    H5Pset_filter
    H5Pset_layout
    H5Pset_shuffle
    H5Pset_szip

Dataset transfer properties:

    H5Pset_btree_ratios
    H5Pset_buffer
    H5Pset_dxpl_mpio
    H5Pset_hyper_cache
    H5Pset_preserve

File access properties:

    H5Pset_alignment
    H5Pset_cache
    H5Pset_fapl_mpio
    H5Pset_fclose_degree
    H5Pset_gc_references
    H5Pset_meta_block_size
    H5Pset_small_data_block_size
    H5Pset_sieve_buf_size

File creation properties:

    H5Pset_istore_k
    H5Pset_sizes
    H5Pset_sym_k
    H5Pset_userblock

Notes

(1)		All processes must participate only if this is the last reference to the file identifier.
(2)		All processes must participate only if all file identifiers for a file have been closed and this is the last outstanding object identifier.
(3)		Because raw data for an attribute is cached locally, all processes must participate in order to guarantee that future `H5Aread` calls return correct results on all processes.
(4)		All processes must participate only if the datatype is for a committed datatype, all the file identifiers for the file have been closed, and this is the last outstanding object identifier.
(5)		All processes must participate only if the number of chunks in the dataset actually changes.
(6)		All processes must use the same datatype, dataspace, and creation properties.
(7)		This function may be called independently if the object identifier does not refer to an object that was collectively opened.
(8)		All processes must use the same creation properties.
(9)		All processes must use the same access properties.
(10)		All processes must use the same dataspace dimensions
(A)		Available only in the HDF5 Release 1.6.x series or later versions of the library.

HDF5 documents and links Introduction to HDF5 HDF5 User Guide	And in this document, the HDF5 Reference Manual H5 H5A H5D H5E H5F H5G H5I H5P H5R H5S H5T H5Z Tools Datatypes Collective Calls in Parallel
(PDF of complete manual formatted as print volume)

The HDF Group Help Desk: