Advanced HDF5 Features

HDF5 Advanced Topics
Neil Fortner
The HDF Group
The 14th HDF and HDF-EOS Workshop
September 28-30, 2010

Sep. 28-30, 2010

HDF/HDF-EOS Workshop XIV

1

Outline
• Overview of HDF5 datatypes
• Partial I/O in HDF5
• Chunking and compression

Sep. 28-30, 2010


2

HDF5 Datatypes
Quick overview of the most
difficult topics

Sep. 28-30, 2010


3

An HDF5 Datatype is…
• A description of dataset element type
• Grouped into “classes”:
•
•
•
•
•
•

Atomic – integers, floating-point values
Enumerated
Compound – like C structs
Array
Opaque
References
• Object – similar to soft link
• Region – similar to soft link to dataset + selection

• Variable-length
• Strings – fixed and variable-length
• Sequences – similar to Standard C++ vector class
Sep. 28-30, 2010


4

HDF5 Datatypes
• HDF5 has a rich set of pre-defined datatypes and
supports the creation of an unlimited variety of
complex user-defined datatypes.
• Self-describing:
• Datatype definitions are stored in the HDF5 file
with the data.
• Datatype definitions include information such as
byte order (endianness), size, and floating point
representation to fully describe how the data is
stored and to insure portability across platforms.

Sep. 28-30, 2010


5

Datatype Conversion
• Datatypes that are compatible, but not identical
are converted automatically when I/O is
performed
• Compatible datatypes:
• All atomic datatypes are compatible
• Identically structured array, variable-length and
compound datatypes whose base type or fields are
compatible
• Enumerated datatype values on a “by name” basis

• Make datatypes identical for best performance

Sep. 28-30, 2010


6

Datatype Conversion Example
Array of integers on IA32 platform
Native integer is little-endian, 4 bytes

Array of integers on SPARC64 platform
Native integer is big-endian, 8 bytes

H5T_NATIVE_INT

H5T_NATIVE_INT

Little-endian 4 bytes integer

H5Dwrite

H5Dread
H5Dwrite
H5T_STD_I32LE

Sep. 28-30, 2010


VAX G-floating

7

Datatype Conversion
Datatype of data on disk
dataset = H5Dcreate(file, DATASETNAME, H5T_STD_I64BE,
space, H5P_DEFAULT, H5P_DEFAULT);

Datatype of data in memory buffer
H5Dwrite(dataset, H5T_NATIVE_INT, H5S_ALL, H5S_ALL,
H5P_DEFAULT, buf);
H5Dwrite(dataset, H5T_NATIVE_DOUBLE, H5S_ALL, H5S_ALL,
H5P_DEFAULT, buf);

Sep. 28-30, 2010


8

Storing Records with HDF5

Sep. 28-30, 2010


9

HDF5 Compound Datatypes
• Compound types
• Comparable to C structs
• Members can be any datatype
• Can write/read by a single field or a set of fields
• Not all data filters can be applied
(shuffling, SZIP)

Sep. 28-30, 2010


10

Creating and Writing Compound Dataset
h5_compound.c example

typedef struct s1_t {
int a;
float b;
double c;
} s1_t;
s1_t

Sep. 28-30, 2010

s1[LENGTH];


11

/* Create datatype in memory. */
s1_tid = H5Tcreate(H5T_COMPOUND, sizeof(s1_t));
H5Tinsert(s1_tid, "a_name", HOFFSET(s1_t, a),
H5T_NATIVE_INT);
H5Tinsert(s1_tid, "c_name", HOFFSET(s1_t, c),
H5T_NATIVE_DOUBLE);
H5Tinsert(s1_tid, "b_name", HOFFSET(s1_t, b),
H5T_NATIVE_FLOAT);

Note:
• Use HOFFSET macro instead of calculating offset by hand.
• Order of H5Tinsert calls is not important if HOFFSET is used.

Sep. 28-30, 2010


12

/* Create dataset and write data */
dataset = H5Dcreate(file, DATASETNAME, s1_tid, space,
H5P_DEFAULT, H5P_DEFAULT);
status = H5Dwrite(dataset, s1_tid, H5S_ALL, H5S_ALL,
H5P_DEFAULT, s1);
Note:
• In this example memory and file datatypes are the same.
• Type is not packed.
• Use H5Tpack to save space in the file.
status = H5Tpack(s1_tid);
status = H5Dcreate(file, DATASETNAME, s1_tid, space,
H5P_DEFAULT, H5P_DEFAULT);

Sep. 28-30, 2010


13

Reading Compound Dataset
/* Create datatype in memory and read data. */
dataset
= H5Dopen(file, DATASETNAME, H5P_DEFAULT);
s2_tid
= H5Dget_type(dataset);
mem_tid
= H5Tget_native_type(s2_tid);
buf = malloc(H5Tget_size(mem_tid)*number_of_elements);
status
= H5Dread(dataset, mem_tid, H5S_ALL,
H5S_ALL, H5P_DEFAULT, buf);
Note:
• We could construct memory type as we did in writing example.
• For general applications we need to discover the type in the
file, find out corresponding memory type, allocate space and do
read.

Sep. 28-30, 2010


14

Reading Compound Dataset by Fields
typedef struct s2_t {
double c;
int
a;
} s2_t;
s2_t s2[LENGTH];
…
s2_tid = H5Tcreate (H5T_COMPOUND, sizeof(s2_t));
H5Tinsert(s2_tid, "c_name", HOFFSET(s2_t, c),
H5T_NATIVE_DOUBLE);
H5Tinsert(s2_tid, “a_name", HOFFSET(s2_t, a),
H5T_NATIVE_INT);
…
status = H5Dread(dataset, s2_tid, H5S_ALL,
H5S_ALL, H5P_DEFAULT, s2);

Sep. 28-30, 2010


15

Table Example
a_name b_name c_name
(integer) (float) (double)
0
0.
1.0000
1
2
3

1.
4.
9.

0.5000
0.3333
0.2500

4
5
6
7
8
9

16.
25.
36.
49.
64.
81.

0.2000
0.1667
0.1429
0.1250
0.1111
0.1000

Sep. 28-30, 2010

Multiple ways to store a table
• Dataset for each field
• Dataset with compound datatype
• If all fields have the same type:
◦ 2-dim array
◦ 1-dim array of array datatype
• Continued…

Choose to achieve your goal!
•
•
•
•
•

Storage overhead?
Do I always read all fields?
Do I read some fields more often?
Do I want to use compression?
Do I want to access some records?


16

Storing Variable Length
Data with HDF5

Sep. 28-30, 2010


17

HDF5 Fixed and Variable Length Array Storage

•Data
•Data

Time
•Data
•Data

•Data
•Data

Time
•Data

•Data
•Data

Sep. 28-30, 2010


18

Storing Variable Length Data in HDF5
• Each element is represented by C structure
typedef struct {
size_t length;
void
*p;
} hvl_t;

• Base type can be any HDF5 type
H5Tvlen_create(base_type)

Sep. 28-30, 2010


19

Example
hvl_t

data[LENGTH];

for(i=0; i<LENGTH; i++) {
data[i].p = malloc((i+1)*sizeof(unsigned int));
data[i].len = i+1;
}
tvl = H5Tvlen_create (H5T_NATIVE_UINT);

data[0].p

•Data
•Data
•Data
•Data

data[4].len
Sep. 28-30, 2010

•Data


20

Reading HDF5 Variable Length Array
• HDF5 library allocates memory to read data in
• Application only needs to allocate array of hvl_t
elements (pointers and lengths)
• Application must reclaim memory for data read in
hvl_t rdata[LENGTH];
/* Create the memory vlen type */
tvl = H5Tvlen_create(H5T_NATIVE_INT);
ret = H5Dread(dataset, tvl, H5S_ALL, H5S_ALL,
H5P_DEFAULT, rdata);
/* Reclaim the read VL data */
H5Dvlen_reclaim(tvl, H5S_ALL, H5P_DEFAULT,rdata);

Sep. 28-30, 2010


21

Variable Length vs. Array
• Pros of variable length datatypes vs. arrays:
• Uses less space if compression unavailable
• Automatically stores length of data
• No maximum size
• Size of an array is its effective maximum size

• Cons of variable length datatypes vs. arrays:
• Substantial performance overhead
• Each element a “pointer” to piece of metadata

• Variable length data cannot be compressed
• Unused space in arrays can be “compressed away”

• Must be 1-dimensional
Sep. 28-30, 2010


22

Storing Strings in HDF5

Sep. 28-30, 2010


23

Storing Strings in HDF5
• Array of characters (Array datatype or extra dimension in
dataset)
• Quick access to each character
• Extra work to access and interpret each string
• Fixed length
string_id = H5Tcopy(H5T_C_S1);
H5Tset_size(string_id, size);

• Wasted space in shorter strings
• Can be compressed
• Variable length
string_id = H5Tcopy(H5T_C_S1);
H5Tset_size(string_id, H5T_VARIABLE);

• Overhead as for all VL datatypes
• Compression will not be applied to actual data

Sep. 28-30, 2010


24

HDF5 Reference Datatypes

Sep. 28-30, 2010


25

Reference Datatypes
• Object Reference
• Pointer to an object in a file
• Predefined datatype H5T_STD_REG_OBJ
• Dataset Region Reference
• Pointer to a dataset + dataspace selection
• Predefined datatype
H5T_STD_REF_DSETREG

Sep. 28-30, 2010


26

Saving Selected Region in a File
Need to select and access the same
elements of a dataset

Sep. 28-30, 2010


27

Reference to Dataset Region
REF_REG.h5
Root

Matrix

Region References

1 1 2 3 3 4 5 5 6
1 2 2 3 4 4 5 6 6

Sep. 28-30, 2010


28

Working with subsets

Sep. 28-30, 2010


30

Collect data one way ….
Array of images (3D)

Sep. 28-30, 2010


31

Display data another way …

Stitched image (2D array)

Sep. 28-30, 2010


32

Data is too big to read….

Sep. 28-30, 2010


33

HDF5 Library Features
• HDF5 Library provides capabilities to
• Describe subsets of data and perform write/read
operations on subsets
• Hyperslab selections and partial I/O

• Store descriptions of the data subsets in a file
• Object references
• Region references

• Use efficient storage mechanism to achieve good
performance while writing/reading subsets of data
• Chunking, compression

Sep. 28-30, 2010


34

Partial I/O in HDF5

Sep. 28-30, 2010


35

How to Describe a Subset in HDF5?
• Before writing and reading a subset of data
one has to describe it to the HDF5 Library.
• HDF5 APIs and documentation refer to a
subset as a “selection” or “hyperslab
selection”.
• If specified, HDF5 Library will perform I/O on a
selection only and not on all elements of a
dataset.

Sep. 28-30, 2010


36

Types of Selections in HDF5
• Two types of selections
• Hyperslab selection
• Regular hyperslab
• Simple hyperslab
• Result of set operations on hyperslabs
(union, difference, …)

• Point selection

• Hyperslab selection is especially important for
doing parallel I/O in HDF5 (See Parallel HDF5
Tutorial)

Sep. 28-30, 2010


37

Regular Hyperslab

Collection of regularly spaced equal size blocks

Sep. 28-30, 2010


38

Simple Hyperslab

Contiguous subset or sub-array

Sep. 28-30, 2010


39

Hyperslab Selection

Result of union operation on three simple hyperslabs

Sep. 28-30, 2010


40

Hyperslab Description
• Start - starting location of a hyperslab (1,1)
• Stride - number of elements that separate each
block (3,2)
• Count - number of blocks (2,6)
• Block - block size (2,1)
• Everything is “measured” in number of elements

Sep. 28-30, 2010


41

Simple Hyperslab Description
• Two ways to describe a simple hyperslab
• As several blocks
• Stride – (1,1)
• Count – (4,6)
• Block – (1,1)

• As one block
• Stride – (1,1)
• Count – (1,1)
• Block – (4,6)

No performance penalty for
one way or another
Sep. 28-30, 2010


42

H5Sselect_hyperslab Function

space_id Identifier of dataspace
op
Selection operator
H5S_SELECT_SET or H5S_SELECT_OR
start
Array with starting coordinates of hyperslab
stride
Array specifying which positions along a dimension
to select
count
Array specifying how many blocks to select from the
dataspace, in each dimension
block
Array specifying size of element block
(NULL indicates a block size of a single element in
a dimension)

Sep. 28-30, 2010


43

Reading/Writing Selections
Programming model for reading from a dataset in
a file
1. Open a dataset.
2. Get file dataspace handle of the dataset and specify
subset to read from.
a. H5Dget_space returns file dataspace handle
a.

File dataspace describes array stored in a file (number of
dimensions and their sizes).

b. H5Sselect_hyperslab selects elements of the array
that participate in I/O operation.

3. Allocate data buffer of an appropriate shape and size

Sep. 28-30, 2010


44

Reading/Writing Selections
Programming model (continued)
4. Create a memory dataspace and specify subset to write
to.
1.
2.

Memory dataspace describes data buffer (its rank and
dimension sizes).
Use H5Screate_simple function to create memory
dataspace.

Use H5Sselect_hyperslab to select elements of the data
buffer that participate in I/O operation.
Issue H5Dread or H5Dwrite to move the data between

3.

5.

file and memory buffer.
6. Close file dataspace and memory dataspace when
done.
Sep. 28-30, 2010


45

Example : Reading Two Rows
1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

-1

-1

-1

Data in a file
4x6 matrix

Buffer in memory
1-dim array of length 14
-1

-1

Sep. 28-30, 2010

-1

-1

-1

-1

-1


-1

-1
46

-1

-1

Example: Reading Two Rows
1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

start
count
block
stride

24

filespace = H5Dget_space (dataset);
H5Sselect_hyperslab (filespace, H5S_SELECT_SET,
start, NULL, count, NULL)

Sep. 28-30, 2010


47

=
=
=
=

{1,0}
{2,6}
{1,1}
{1,1}

start[1] = {1}
count[1] = {12}
dim[1]
= {14}

-1

-1

-1

-1

-1

-1

-1

-1

-1

-1

-1

-1

memspace = H5Screate_simple(1, dim, NULL);
H5Sselect_hyperslab (memspace, H5S_SELECT_SET,
start, NULL, count, NULL)

Sep. 28-30, 2010


48

-1

-1

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

H5Dread (…, …, memspace, filespace, …, …);

-1

7

Sep. 28-30, 2010

8

9

10 11 12 13 14 15 16 17 18 -1

49

Things to Remember
• Number of elements selected in a file and in a
memory buffer must be the same
• H5Sget_select_npoints returns number of
selected elements in a hyperslab selection

• HDF5 partial I/O is tuned to move data between
selections that have the same dimensionality;
avoid choosing subsets that have different ranks
(as in example above)
• Allocate a buffer of an appropriate size when
reading data; use H5Tget_native_type and
H5Tget_size to get the correct size of the data
element in memory.
Sep. 28-30, 2010


50

Chunking in HDF5

Sep. 28-30, 2010


51

HDF5 Dataset

Metadata

Dataset data

Dataspace
Rank Dimensions
3

Dim_1 = 4
Dim_2 = 5
Dim_3 = 7

Datatype
IEEE 32-bit float

Attributes
Storage info

Time = 32.4

Chunked

Pressure = 987

Compressed

Temp = 56

Sep. 28-30, 2010


52

Contiguous storage layout
• Metadata header separate from dataset data
• Data stored in one contiguous block in HDF5 file
Metadata cache
Dataset header
………….
Datatype
Dataspace
………….
Attributes
…

Dataset data

Application memory

File

Sep. 28-30, 2010

Dataset data


53

What is HDF5 Chunking?
• Data is stored in chunks of predefined size
• Two-dimensional instance may be referred to as
data tiling
• HDF5 library usually writes/reads the whole chunk

Contiguous

Sep. 28-30, 2010


Chunked

54

What is HDF5 Chunking?
• Dataset data is divided into equally sized blocks (chunks).
• Each chunk is stored separately as a contiguous block in
HDF5 file.
Metadata cache

Dataset data

Dataset header
………….
Datatype
Dataspace
………….
Attributes
…

File
Sep. 28-30, 2010

A

B

C

D

Chunk
index

Application memory

header

Chunk
index

A


C

D

B
55

Why HDF5 Chunking?
• Chunking is required for several HDF5 features
• Enabling compression and other filters like
checksum
• Extendible datasets

Sep. 28-30, 2010


56

Why HDF5 Chunking?
• If used appropriately chunking improves partial
I/O for big datasets

Only two chunks are involved in I/O

Sep. 28-30, 2010


57

Creating Chunked Dataset

1.
2.
3.

Create a dataset creation property list.
Set property list to use chunked storage layout.
Create dataset with the above property list.
dcpl_id = H5Pcreate(H5P_DATASET_CREATE);
rank = 2;
ch_dims[0] = 100;
ch_dims[1] = 200;
H5Pset_chunk(dcpl_id, rank, ch_dims);
dset_id = H5Dcreate (…, dcpl_id);
H5Pclose(dcpl_id);

Sep. 28-30, 2010


58

• Things to remember:
• Chunk always has the same rank as a dataset
• Chunk’s dimensions do not need to be factors
of dataset’s dimensions
• Caution: May cause more I/O than desired
(see white portions of the chunks below)

Sep. 28-30, 2010


59

• Chunk size cannot be changed after the dataset is
created
• Do not make chunk sizes too small (e.g. 1x1)!
• Metadata overhead for each chunk (file space)
• Each chunk is read individually
• Many small reads inefficient

Sep. 28-30, 2010


60

Writing or Reading Chunked Dataset
1.
2.

Chunking mechanism is transparent to application.
Use the same set of operation as for contiguous
dataset, for example,
H5Dopen(…);
H5Sselect_hyperslab (…);
H5Dread(…);

3.

Selections do not need to coincide precisely with the
chunks boundaries.

Sep. 28-30, 2010


61

HDF5 Chunking and compression
•

Chunking is required for compression and
other filters
HDF5 filters modify data during I/O operations
Filters provided by HDF5:

•
•
•
•
•

•

Checksum (H5Pset_fletcher32)
Data transformation (in 1.8.*)
Shuffling filter (H5Pset_shuffle)

Compression (also called filters) in HDF5
•
•
•
•

Sep. 28-30, 2010

Scale + offset (in 1.8.*) (H5Pset_scaleoffset)
N-bit (in 1.8.*) (H5Pset_nbit)
GZIP (deflate) (H5Pset_deflate)
SZIP (H5Pset_szip)

62

HDF5 Third-Party Filters
• Compression methods supported by
HDF5 User’s community
http://wiki.hdfgroup.org/Community-Support-for-HDF5
•
LZO lossless compression (PyTables)
•
BZIP2 lossless compression (PyTables)
•
BLOSC lossless compression (PyTables)
•
LZF lossless compression H5Py

Sep. 28-30, 2010


63

Creating Compressed Dataset
1.
2.
3.
4.

Create a dataset creation property list
Set property list to use chunked storage layout
Set property list to use filters
Create dataset with the above property list

dcpl_id = H5Pcreate(H5P_DATASET_CREATE);
rank = 2;
ch_dims[0] = 100;
ch_dims[1] = 100;
H5Pset_chunk(dcpl_id, rank, ch_dims);
H5Pset_deflate(dcpl_id, 9);
dset_id = H5Dcreate (…, dcpl_id);
H5Pclose(dcpl_id);

Sep. 28-30, 2010


64

Performance Issues
or
What everyone needs to know
about chunking and the chunk
cache

Sep. 28-30, 2010


65

Accessing a row in contiguous dataset

One seek is needed to find the starting location of row of data.
Data is read/written using one disk access.

Sep. 28-30, 2010


66

Accessing a row in chunked dataset

Five seeks is needed to find each chunk. Data is read/written
using five disk accesses. Chunking storage is less efficient
than contiguous storage.
Sep. 28-30, 2010


67

Quiz time
• How might I improve this situation, if it is
common to access my data in this way?

Sep. 28-30, 2010


68

Accessing data in contiguous dataset

M rows

M seeks are needed to find the starting location of the element.
Data is read/written using M disk accesses. Performance may be
very bad.

Sep. 28-30, 2010


69

Motivation for chunking storage

M rows

Two seeks are needed to find two chunks. Data is
read/written using two disk accesses. For this pattern
chunking helps with I/O performance.

Sep. 28-30, 2010


70

Motivation for chunk cache
A

B

H5Dwrite
H5Dwrite

Selection shown is written by two H5Dwrite calls (one for
each row).
Chunks A and B are accessed twice (one time for each
row). If both chunks fit into cache, only two I/O accesses
needed to write the shown selections.
Sep. 28-30, 2010


71

Motivation for chunk cache
A

B

H5Dwrite
H5Dwrite

Question: What happens if there is a space for only one
chunk at a time?

Sep. 28-30, 2010


72

Advanced Exercise
•
•
•
•

Write data to a dataset
Dataset is 512x2048, 4-byte native integers
Chunks are 256x128: 128KB each, 2MB rows
Write by rows

Sep. 28-30, 2010


73

Advanced Exercise
• Very slow performance
• What is going wrong?
• Chunk cache is only 1MB by default
Read into cache

Sep. 28-30, 2010


74

Advanced Exercise
Write to disk

Sep. 28-30, 2010

Read into cache


75

Advanced Exercise
Write to disk

Sep. 28-30, 2010

Read into cache


76

Advanced Exercise
Write to disk

Sep. 28-30, 2010

Read into cache


77

Advanced Exercise
Write to disk

Sep. 28-30, 2010


Read into cache

78

Advanced Exercise
Write to disk

Sep. 28-30, 2010


Read into cache

79

Advanced Exercise
Read into cache

Sep. 28-30, 2010

Write to disk


80

Advanced Exercise
Read into cache

Sep. 28-30, 2010

Write to disk


81

Exercise 1
• Improve performance by changing only chunk
size
Access pattern is fixed, limited memory
• One solution: 64x2048 chunks
• Row of chunks fits in cache

Sep. 28-30, 2010


82

Exercise 2
• Improve performance by changing only access
pattern
• File already exists, cannot change chunk size

• One solution: Access by chunk
• Each selection fits in cache, contiguous on disk

Sep. 28-30, 2010


83

Exercise 3
• Improve performance while not changing chunk
size or access pattern
• No memory limitation

• One solution: Chunk cache set to size of row of
chunks

Sep. 28-30, 2010


84

Exercise 4
• Improve performance while not changing chunk
size or access pattern
• Chunk cache size can be set to max. 1MB
• One solution: Disable chunk cache
• Avoids repeatedly reading/writing whole chunks

Sep. 28-30, 2010


85

More Information
• More detailed information on chunking and the
chunk cache can be found in the draft “Chunking
in HDF5” document at:
http://www.hdfgroup.org/HDF5/doc/_topic/Chunking

Sep. 28-30, 2010


86

Thank You!

Sep. 28-30, 2010


87

Acknowledgements
This work was supported by cooperative agreement
number NNX08AO77A from the National
Aeronautics and Space Administration (NASA).
Any opinions, findings, conclusions, or
recommendations expressed in this material are
those of the author[s] and do not necessarily reflect
the views of the National Aeronautics and Space
Administration.

Sep. 28-30, 2010


88

Questions/comments?

Sep. 28-30, 2010


89

Advanced HDF5 Features

More Related Content

Similar to Advanced HDF5 Features (20)

More from The HDF-EOS Tools and Information Center (20)

Recently uploaded (20)

Advanced HDF5 Features