HDF5 documents and links 
Introduction to HDF5 
HDF5 Reference Manual 
HDF5 User's Guide for Release 1.6 
And in this document, the HDF5 User's Guide from Release 1.4.5:    
Files   Datasets   Datatypes   Dataspaces   Groups  
References   Attributes   Property Lists   Error Handling  
Filters   Caching   Chunking   Mounting Files  
Performance   Debugging   Environment   DDL  

The Dataset Interface (H5D)

1. Introduction

The purpose of the dataset interface is to provide a mechanism to describe properties of datasets and to transfer data between memory and disk. A dataset is composed of a collection of raw data points and four classes of meta data to describe the data points. The interface is hopefully designed in such a way as to allow new features to be added without disrupting current applications that use the dataset interface.

The four classes of meta data are:

Constant Meta Data
Meta data that is created when the dataset is created and exists unchanged for the life of the dataset. For instance, the datatype of stored array elements is defined when the dataset is created and cannot be subsequently changed.
Persistent Meta Data
Meta data that is an integral and permanent part of a dataset but can change over time. For instance, the size in any dimension can increase over time if such an increase is allowed when the dataset was created.
Memory Meta Data
Meta data that exists to describe how raw data is organized in the application's memory space. For instance, the data type of elements in an application array might not be the same as the datatype of those elements as stored in the HDF5 file.
Transport Meta Data
Meta data that is used only during the transfer of raw data from one location to another. For instance, the number of processes participating in a collective I/O request or hints to the library to control caching of raw data.

Each of these classes of meta data is handled differently by the library although the same API might be used to create them. For instance, the datatype exists as constant meta data and as memory meta data; the same API (the H5T API) is used to manipulate both pieces of meta data but they're handled by the dataset API (the H5D API) in different manners.

2. Storage Layout Properties

The dataset API partitions these terms on three orthogonal axes (layout, compression, and external storage) and uses a dataset creation property list to hold the various settings and pass them through the dataset interface. This is similar to the way HDF5 files are created with a file creation property list. A dataset creation property list is always derived from the default dataset creation property list (use H5Pcreate() to get a copy of the default property list) by modifying properties with various H5Pset_property() functions.

herr_t H5Pset_layout (hid_t plist_id, H5D_layout_t layout)
The storage layout is a piece of constant meta data that describes what method the library uses to organize the raw data on disk. The default layout is contiguous storage.

H5D_COMPACT    (Not yet implemented.)
The raw data is presumably small and can be stored directly in the object header. Such data is non-extendible, non-compressible, non-sparse, and cannot be stored externally. Most of these restrictions are arbitrary but are enforced because of the small size of the raw data. Storing data in this format eliminates the disk seek/read request normally necessary to read raw data.

The raw data is large, non-extendible, non-compressible, non-sparse, and can be stored externally. This is the default value for the layout property. The term large means that it may not be possible to hold the entire dataset in memory. The non-compressibility is a side effect of the data being large, contiguous, and fixed-size at the physical level, which could cause partial I/O requests to be extremely expensive if compression were allowed.

The raw data is large and can be extended in any dimension at any time (provided the data space also allows the extension). It may be sparse at the chunk level (each chunk is non-sparse, but there might only be a few chunks) and each chunk can be compressed and/or stored externally. A dataset is partitioned into chunks so each chunk is the same logical size. The chunks are indexed by a B-tree and are allocated on demand (although it might be useful to be able to preallocate storage for parts of a chunked array to reduce contention for the B-tree in a parallel environment). The chunk size must be defined with H5Pset_chunk().

Other layout types may be defined later without breaking existing code. However, to be able to correctly read or modify data stored with one of these new layouts, the application will need to be linked with a new version of the library. This happens automatically on systems with dynamic linking.

Once the general layout is defined, the user can define properties of that layout. Currently, the only layout that has user-settable properties is the H5D_CHUNKED layout, which needs to know the dimensionality and chunk size.

herr_t H5Pset_chunk (hid_t plist_id, int ndims, hsize_t dim[])
This function defines the logical size of a chunk for chunked layout. If the layout property is set to H5D_CHUNKED and the chunk size is set to dim. The number of elements in the dim array is the dimensionality, ndims. One need not call H5Dset_layout() when using this function since the chunked layout is implied.

Example: Chunked Storage

This example shows how a two-dimensional dataset is partitioned into chunks. The library can manage file memory by moving the chunks around, and each chunk could be compressed. The chunks are allocated in the file on demand when data is written to the chunk.

Chunked Storage

size_t hsize[2] = {1000, 1000};
plist = H5Pcreate (H5P_DATASET_CREATE);
H5Pset_chunk (plist, 2, size);

Although it is most efficient if I/O requests are aligned on chunk boundaries, this is not a constraint. The application can perform I/O on any set of data points as long as the set can be described by the data space. The set on which I/O is performed is called the selection.

3. Compression Properties

Chunked data storage (see H5Pset_chunk) allows data compression as defined by the function H5Pset_deflate.

herr_t H5Pset_deflate (hid_t plist_id, int level)
int H5Pget_deflate (hid_t plist_id)
These functions set or query the deflate level of dataset creation property list plist_id. The H5Pset_deflate() sets the compression method to H5Z_DEFLATE and sets the compression level to some integer between one and nine (inclusive). One results in the fastest compression while nine results in the best compression ratio. The default value is six if H5Pset_deflate() isn't called. The H5Pget_deflate() returns the compression level for the deflate method, or negative if the method is not the deflate method.

4. External Storage Properties

Some storage formats may allow storage of data across a set of non-HDF5 files. Currently, only the H5D_CONTIGUOUS storage format allows external storage. A set segments (offsets and sizes) in one or more files is defined as an external file list, or EFL, and the contiguous logical addresses of the data storage are mapped onto these segments.

herr_t H5Pset_external (hid_t plist, const char *name, off_t offset, hsize_t size)
This function adds a new segment to the end of the external file list of the specified dataset creation property list. The segment begins a byte offset of file name and continues for size bytes. The space represented by this segment is adjacent to the space already represented by the external file list. The last segment in a file list may have the size H5F_UNLIMITED, in which case the external file may be of unlimited size and no more files can be added to the external files list.

int H5Pget_external_count (hid_t plist)
Calling this function returns the number of segments in an external file list. If the dataset creation property list has no external data then zero is returned.

herr_t H5Pget_external (hid_t plist, unsigned idx, size_t name_size, char *name, off_t *offset, hsize_t *size)
This is the counterpart for the H5Pset_external() function. Given a dataset creation property list and a zero-based index into that list, the file name, byte offset, and segment size are returned through non-null arguments. At most name_size characters are copied into the name argument which is not null terminated if the file name is longer than the supplied name buffer (this is similar to strncpy()).

Example: Multiple Segments

This example shows how a contiguous, one-dimensional dataset is partitioned into three parts and each of those parts is stored in a segment of an external file. The top rectangle represents the logical address space of the dataset while the bottom rectangle represents an external file.

Multiple Segments

plist = H5Pcreate (H5P_DATASET_CREATE);
H5Pset_external (plist, "velocity.data", 3000, 1000);
H5Pset_external (plist, "velocity.data", 0, 2500);
H5Pset_external (plist, "velocity.data", 4500, 1500);

One should note that the segments are defined in order of the logical addresses they represent, not their order within the external file. It would also have been possible to put the segments in separate files. Care should be taken when setting up segments in a single file since the library doesn't automatically check for segments that overlap.

Example: Multi-Dimensional

This example shows how a contiguous, two-dimensional dataset is partitioned into three parts and each of those parts is stored in a separate external file. The top rectangle represents the logical address space of the dataset while the bottom rectangles represent external files.

Multiple Dimensions

plist = H5Pcreate (H5P_DATASET_CREATE);
H5Pset_external (plist, "scan1.data", 0, 24);
H5Pset_external (plist, "scan2.data", 0, 24);
H5Pset_external (plist, "scan3.data", 0, 16);

The library maps the multi-dimensional array onto a linear address space like normal, and then maps that address space into the segments defined in the external file list.

The segments of an external file can exist beyond the end of the file. The library reads that part of a segment as zeros. When writing to a segment that exists beyond the end of a file, the file is automatically extended. Using this feature, one can create a segment (or set of segments) which is larger than the current size of the dataset, which allows to dataset to be extended at a future time (provided the data space also allows the extension).

All referenced external data files must exist before performing raw data I/O on the dataset. This is normally not a problem since those files are being managed directly by the application, or indirectly through some other library.

5. Datatype

Raw data has a constant datatype which describes the datatype of the raw data stored in the file, and a memory datatype that describes the datatype stored in application memory. Both data types are manipulated with the H5T API.

The constant file datatype is associated with the dataset when the dataset is created in a manner described below. Once assigned, the constant datatype can never be changed.

The memory datatype is specified when data is transferred to/from application memory. In the name of data sharability, the memory datatype must be specified, but can be the same type identifier as the constant datatype.

During dataset I/O operations, the library translates the raw data from the constant datatype to the memory datatype or vice versa. Structured datatypes include member offsets to allow reordering of struct members and/or selection of a subset of members and array datatypes include index permutation information to allow things like transpose operations (the prototype does not support array reordering) Permutations are relative to some extrinsic descritpion of the dataset.

6. Data Space

The dataspace of a dataset defines the number of dimensions and the size of each dimension and is manipulated with the H5S API. The simple dataspace consists of maximum dimension sizes and actual dimension sizes, which are usually the same. However, maximum dimension sizes can be the constant H5D_UNLIMITED in which case the actual dimension size can be incremented with calls to H5Dextend(). The maximium dimension sizes are constant meta data while the actual dimension sizes are persistent meta data. Initial actual dimension sizes are supplied at the same time as the maximum dimension sizes when the dataset is created.

The dataspace can also be used to define partial I/O operations. Since I/O operations have two end-points, the raw data transfer functions take two data space arguments: one which describes the application memory data space or subset thereof and another which describes the file data space or subset thereof.

7. Setting Constant or Persistent Properties

Each dataset has a set of constant and persistent properties which describe the layout method, pre-compression transformation, compression method, datatype, external storage, and data space. The constant properties are set as described above in a dataset creation property list whose identifier is passed to H5Dcreate().

hid_t H5Dcreate (hid_t file_id, const char *name, hid_t type_id, hid_t space_id, hid_t create_plist_id)
A dataset is created by calling H5Dcreate with a file identifier, a dataset name, a datatype, a dataspace, and constant properties. The datatype and dataspace are the type and space of the dataset as it will exist in the file, which may be different than in application memory. Dataset names within a group must be unique: H5Dcreate returns an error if a dataset with the name specified in name already exists at the location specified in file_id. The create_plist_id is a H5P_DATASET_CREATE property list created with H5Pcreate() and initialized with the various functions described above. H5Dcreate() returns a dataset handle for success or negative for failure. The handle should eventually be closed by calling H5Dclose() to release resources it uses.

hid_t H5Dopen (hid_t file_id, const char *name)
An existing dataset can be opened for access by calling this function. A dataset handle is returned for success or a negative value is returned for failure. The handle should eventually be closed by calling H5Dclose() to release resources it uses.

herr_t H5Dclose (hid_t dataset_id)
This function closes a dataset handle and releases all resources it might have been using. The handle should not be used in subsequent calls to the library.

herr_t H5Dextend (hid_t dataset_id, hsize_t dim[])
This function extends a dataset by increasing the size in one or more dimensions. Not all datasets can be extended.

8. Querying Constant or Persistent Properties

Constant or persistent properties can be queried with a set of three functions. Each function returns an identifier for a copy of the requested properties. The identifier can be passed to various functions which modify the underlying object to derive a new object; the original dataset is completely unchanged. The return values from these functions should be properly destroyed when no longer needed.

hid_t H5Dget_type (hid_t dataset_id)
Returns an identifier for a copy of the dataset permanent datatype or negative for failure.
hid_t H5Dget_space (hid_t dataset_id)
Returns an identifier for a copy of the dataset permanent data space, which also contains information about the current size of the dataset if the data set is extendable with H5Dextend().
hid_t H5Dget_create_plist (hid_t dataset_id)
Returns an identifier for a copy of the dataset creation property list. The new property list is created by examining various permanent properties of the dataset. This is mostly a catch-all for everything but type and space.

9. Setting Memory and Transfer Properties

A dataset also has memory properties which describe memory within the application, and transfer properties that control various aspects of the I/O operations. The memory can have a datatype different than the permanent file datatype (different number types, different struct member offsets, different array element orderings) and can also be a different size (memory is a subset of the permanent dataset elements, or vice versa). The transfer properties might provide caching hints or collective I/O information. Therefore, each I/O operation must specify memory and transfer properties.

The memory properties are specified with type_id and space_id arguments while the transfer properties are specified with the transfer_id property list for the H5Dread() and H5Dwrite() functions (these functions are described below).

herr_t H5Pset_buffer (hid_t xfer_plist, hsize_t max_buf_size, void *tconv_buf, void *bkg_buf)
hsize_t H5Pget_buffer (hid_t xfer_plist, void **tconv_buf, void **bkg_buf)
Sets or retrieves the maximum size in bytes of the temporary buffer used for datatype conversion in the I/O pipeline. An application-defined buffer can also be supplied as the tconv_buf argument, otherwise a buffer will be allocated and freed on demand by the library. A second temporary buffer bkg_buf can also be supplied and should be the same size as the tconv_buf. The default values are 1MB for the maximum buffer size, and null pointers for each buffer indicating that they should be allocated on demand and freed when no longer needed. The H5Pget_buffer() function returns the maximum buffer size or zero on error.

If the maximum size of the temporary I/O pipeline buffers is too small to hold the entire I/O request, then the I/O request will be fragmented and the transfer operation will be strip mined. However, certain restrictions apply to the strip mining. For instance, when performing I/O on a hyperslab of a simple data space the strip mining is in terms of the slowest varying dimension. So if a 100x200x300 hyperslab is requested, the temporary buffer must be large enough to hold a 1x200x300 sub-hyperslab.

To prevent strip mining from happening, the application should use H5Pset_buffer() to set the size of the temporary buffer so it's large enough to hold the entire request.


This example shows how to define a function that sets a dataset transfer property list so that strip mining does not occur. It takes an (optional) dataset transfer property list, a dataset, a data space that describes what data points are being transfered, and a datatype for the data points in memory. It returns a (new) dataset transfer property list with the temporary buffer size set to an appropriate value. The return value should be passed as the fifth argument to H5Dread() or H5Dwrite().

 1 hid_t
 2 disable_strip_mining (hid_t xfer_plist, hid_t dataset,
 3                       hid_t space, hid_t mem_type)
 4 {
 5     hid_t file_type;          /* File datatype */
 6     size_t type_size;         /* Sizeof larger type */
 7     size_t size;              /* Temp buffer size */
 8     hid_t xfer_plist;         /* Return value */
10     file_type = H5Dget_type (dataset);
11     type_size = MAX(H5Tget_size(file_type), H5Tget_size(mem_type));
12     H5Tclose (file_type);
13     size = H5Sget_npoints(space) * type_size;
14     if (xfer_plist<0) xfer_plist = H5Pcreate (H5P_DATASET_XFER);
15     H5Pset_buffer(xfer_plist, size, NULL, NULL);
16     return xfer_plist;
17 }

10. Querying Memory or Transfer Properties

Unlike constant and persistent properties, a dataset cannot be queried for it's memory or transfer properties. Memory properties cannot be queried because the application already stores those properties separate from the buffer that holds the raw data, and the buffer may hold multiple segments from various datasets and thus have more than one set of memory properties. The transfer properties cannot be queried from the dataset because they're associated with the transfer itself and not with the dataset (but one can call H5Pget_property() to query transfer properties from a tempalate).

11. Raw Data I/O

All raw data I/O is accomplished through these functions which take a dataset handle, a memory datatype, a memory data space, a file data space, transfer properties, and an application memory buffer. They translate data between the memory datatype and space and the file datatype and space. The data spaces can be used to describe partial I/O operations.

herr_t H5Dread (hid_t dataset_id, hid_t mem_type_id, hid_t mem_space_id, hid_t file_space_id, hid_t xfer_plist_id, void *buf/*out*/)
Reads raw data from the specified dataset into buf converting from file datatype and space to memory datatype and space.

herr_t H5Dwrite (hid_t dataset_id, hid_t mem_type_id, hid_t mem_space_id, hid_t file_space_id, hid_t xfer_plist_id, const void *buf)
Writes raw data from an application buffer buf to the specified dataset converting from memory datatype and space to file datatype and space.

In the name of sharability, the memory datatype must be supplied. However, it can be the same identifier as was used to create the dataset or as was returned by H5Dget_type(); the library will not implicitly derive memory datatypes from constant datatypes.

For complete reads of the dataset one may supply H5S_ALL as the argument for the file data space. If H5S_ALL is also supplied as the memory data space then no data space conversion is performed. This is a somewhat dangerous situation since the file data space might be different than what the application expects.

12. Examples

The examples in this section illustrate some common dataset practices.

This example shows how to create a dataset which is stored in memory as a two-dimensional array of native double values but is stored in the file in Cray float format using LZ77 compression. The dataset is written to the HDF5 file and then read back as a two-dimensional array of float values.

Example 1

 1 hid_t file, data_space, dataset, properties;
 2 double dd[500][600];
 3 float ff[500][600];
 4 hsize_t dims[2], chunk_size[2];
 6 /* Describe the size of the array */
 7 dims[0] = 500;
 8 dims[1] = 600;
 9 data_space = H5Screate_simple (2, dims);
12 /*
13  * Create a new file using with read/write access,
14  * default file creation properties, and default file
15  * access properties.
16  */
17 file = H5Fcreate ("test.h5", H5F_ACC_RDWR, H5P_DEFAULT,
18                   H5P_DEFAULT);
20 /* 
21  * Set the dataset creation plist to specify that
22  * the raw data is to be partitioned into 100x100 element
23  * chunks and that each chunk is to be compressed with
24  * LZ77.
25  */
26 chunk_size[0] = chunk_size[1] = 100;
27 properties = H5Pcreate (H5P_DATASET_CREATE);
28 H5Pset_chunk (properties, 2, chunk_size);
29 H5Pset_deflate (properties, 9);
31 /*
32  * Create a new dataset within the file.  The datatype
33  * and data space describe the data on disk, which may
34  * be different than the format used in the application's
35  * memory.
36  */
37 dataset = H5Dcreate (file, "dataset", H5T_CRAY_FLOAT,
38                      data_space, properties);
40 /*
41  * Write the array to the file.  The datatype and data
42  * space describe the format of the data in the `dd'
43  * buffer.  The raw data is translated to the format
44  * required on disk defined above.  We use default raw
45  * data transfer properties.
46  */
47 H5Dwrite (dataset, H5T_NATIVE_DOUBLE, H5S_ALL, H5S_ALL,
48           H5P_DEFAULT, dd);
50 /*
51  * Read the array as floats.  This is similar to writing
52  * data except the data flows in the opposite direction.
53  */
54 H5Dread (dataset, H5T_NATIVE_FLOAT, H5S_ALL, H5S_ALL,
55          H5P_DEFAULT, ff);
64 H5Dclose (dataset);
65 H5Sclose (data_space);
66 H5Pclose (properties);
67 H5Fclose (file);

This example uses the file created in Example 1 and reads a hyperslab of the 500x600 file dataset. The hyperslab size is 100x200 and it is located beginning at element <200,200>. We read the hyperslab into an 200x400 array in memory beginning at element <0,0> in memory. Visually, the transfer looks something like this:

Raw Data Transfer

Example 2

 1 hid_t file, mem_space, file_space, dataset;
 2 double dd[200][400];
 3 hsize_t offset[2];
 4 hsize size[2];
 6 /*
 7  * Open an existing file and its dataset.
 8  */
 9 file = H5Fopen ("test.h5", H5F_ACC_RDONLY, H5P_DEFAULT);
10 dataset = H5Dopen (file, "dataset");
12 /*
13  * Describe the file data space.
14  */
15 offset[0] = 200; /*offset of hyperslab in file*/
16 offset[1] = 200;
17 size[0] = 100;   /*size of hyperslab*/
18 size[1] = 200;
19 file_space = H5Dget_space (dataset);
20 H5Sselect_hyperslab (file_space, H5S_SELECT_SET, offset, NULL, size, NULL);
22 /*
23  * Describe the memory data space.
24  */
25 size[0] = 200;  /*size of memory array*/
26 size[1] = 400;
27 mem_space = H5Screate_simple (2, size);
29 offset[0] = 0;  /*offset of hyperslab in memory*/
30 offset[1] = 0;
31 size[0] = 100;  /*size of hyperslab*/
32 size[1] = 200;
33 H5Sselect_hyperslab (mem_space, H5S_SELECT_SET, offset, NULL, size, NULL);
35 /*
36  * Read the dataset.
37  */
38 H5Dread (dataset, H5T_NATIVE_DOUBLE, mem_space,
39          file_space, H5P_DEFAULT, dd);
41 /*
42  * Close/release resources.
43  */
44 H5Dclose (dataset);
45 H5Sclose (mem_space);
46 H5Sclose (file_space);
47 H5Fclose (file);

If the file contains a compound data structure one of whose members is a floating point value (call it "delta") but the application is interested in reading an array of floating point values which are just the "delta" values, then the application should cast the floating point array as a struct with a single "delta" member.

Example 3

 1 hid_t file, dataset, type;
 2 double delta[200];
 4 /*
 5  * Open an existing file and its dataset.
 6  */
 7 file = H5Fopen ("test.h5", H5F_ACC_RDONLY, H5P_DEFAULT);
 8 dataset = H5Dopen (file, "dataset");
10 /*
11  * Describe the memory datatype, a struct with a single
12  * "delta" member.
13  */
14 type = H5Tcreate (H5T_COMPOUND, sizeof(double));
15 H5Tinsert (type, "delta", 0, H5T_NATIVE_DOUBLE);
17 /*
18  * Read the dataset.
19  */
20 H5Dread (dataset, type, H5S_ALL, H5S_ALL,
21          H5P_DEFAULT, dd);
23 /*
24  * Close/release resources.
25  */
26 H5Dclose (dataset);
27 H5Tclose (type);
28 H5Fclose (file);

HDF5 documents and links 
Introduction to HDF5 
HDF5 Reference Manual 
HDF5 User's Guide for Release 1.6 
And in this document, the HDF5 User's Guide from Release 1.4.5:    
Files   Datasets   Datatypes   Dataspaces   Groups  
References   Attributes   Property Lists   Error Handling  
Filters   Caching   Chunking   Mounting Files  
Performance   Debugging   Environment   DDL  

THG Help Desk:
Describes HDF5 Release 1.4.5, February 2003
Last modified: 2 March 2001