HDF4 and HDF5 Performance Preliminary Results

1
HDF4 and HDF5 PerformancePreliminary Results

• Elena Pourmal
• IV HDF-EOS Workshop
• September 19 - 21 2000

2
Why compare?

• HDF5 emerges as a new standard
• proved to be robust
• most of the planned features have been
  implemented in HDF5-1.2.2
• has a lot of new features compared to HDF4
• time for performance study and tuning
• Users move their data and applications to HDF5
• HDF4 is not bad, but has limited capabilities

3
HDF5 HDF4

• Files over 2GB
• Unlimited number of objects
• One data model (multidimensional array of
  structures)
• support
• Thread safe
• Mounting files
• Diversity of datatypes (compound, VL, opaque) and
  operations (create, write, read, delete, shared)
• Native file is portable
• Modifiable I/O pipe-line (registration of
  compression methods)
• Selections (unions and regular blocks)

• Files less than 2GB
• Max limit 20000 of objects
• Different data models for SD, GR, RI, Vdatas
• N/A
• N/A
• N/A
• Only predefined datatypes such as float32, int16,
  char8
• Native file is not portable
• N/A
• Selections (simple regular subsampling)

4
What to compare?(short list of common features)

• File I/O operations
• plain read and write
• hyperslab selections
• regular subsampling
• access to large number of objects
• storage overhead
• Data organization in the file and access to it
• Vdata vs compound datasets
• Chunking, unlimited dimensions, compression

5
Benchmark Environment

• 440-Mhz UltraSPARC i-IIi
• 1G memory
• Sun OS 5.7
• gettimeofday()
• 2 - 550 Mhz Pentium III Xeon
• 1G memory
• RedHat 6.2
• clock()
• each measurement was taken 10 times, average and
  best times were collected

6
Benchmarks

• Writing 1Dim and 2Dim datasets of integers
• Reading 2Dim contiguous hyperslabs of integers
• Reading 2Dim contiguous hyperslabs of integers
  with subsampling
• Reading fixed size hyperslabs of integers from
  different locations in the dataset
• Writing and reading Vdatas and Compound Datasets
• CERES data

7
Writing 1Dim and 2Dim Datasets
8
Writing 1Dim Datasets

• In this test we created one-dimensional arrays of
  integers with sizes varying from 8Kbytes to 8000
  Kbytes in steps of 8Kbytes. We measured the
  average and best times for writing these arrays
  into HDF4 and HDF5 files.
• Test was performed on Solaris platform. Neither
  HDF4 nor HDF5 performed data conversion.

9
Writing 1Dim Datasets
HDF5 performs about 8 times better than
HDF4. System activity affects timing results.
10
Writing 2Dim Datasets

• In this test we created two-dimensional arrays
  with sizes varying from 40 X 40 bytes to 4000 X
  4000 bytes in steps of 40 bytes for each
  dimension. We measured the average and best times
  for writing these arrays into HDF4 and HDF5
  files. The graphs were plotted by averaging the
  values obtained for the same array size, without
  considering the shape of the array.
• Test was performed on Solaris platform. Neither
  HDF4 nor HDF5 performed data conversion.

11
Writing 2Dim Datasets
HDF4 shows nonlinear growth. HDF5 performs about
10 times better than HDF4.
12
Reading 2Dim Contiguous Hyperslabs
13
Reading Contiguous Hyperslabs

• In this test we created a file with 1000 X 1000
  array of integers. Subsequently, we read
  hyperslabs of different sizes starting from a
  fixed position in the array and the measurements
  for read were averaged over 10 runs. HDF5-1.2.2,
  HDF5-1.2.2-patched and HDF5 development libraries
  were tested.
• Test was performed on Solaris platform. Neither
  HDF4 nor HDF5 performed data conversion.

14
Reading Hyperslabs
For hyperslabs gt 1MB, HDF5 becomes more than 3
times slower than HDF4. It also shows nonlinear
growth.
15
Reading Hyperslabs (latest
version of the HDF5 development branch)
For hyperslabs gt 2MB, HDF5 becomes more about 1.5
times slower than HDF4. It still shows nonlinear
growth.
16
Reading contiguous hyperslabs(fixed size)

• In this test, the size of the hyperslab was fixed
  to 100x100 elements. The hyperslab was moved,
  first along the X axis, then along the Y axis,
  and finally along the diagonal and the read
  performance was measured.
• Test was performed on Solaris platform. Neither
  HDF4 nor HDF5 performed data conversion.

17
Reading 100x100 Hyperslabs from Different
Locations
For small hyperslabs HDF5 performs about 3 times
better than HDF4.
18
Reading Hyperslabs with Subsampling
19
Subsampling Hyperslabs

• In this test we created a file with 1000x1000
  array of integers. Subsequently, we read every
  second element of the hyperslabs of different
  sizes starting from a fixed position in the array
  and the measurements for read were averaged over
  10 runs. HDF5-1.2.2, and HDF5 development
  libraries were tested.
• Test was performed on Solaris platform. Neither
  HDF4 nor HDF5 performed data conversion.

20
Reading Each Second Element of the Hyperslabs
HDF5 shows nonlinear growth. HDF4 performs about
3 times for the hyperslabs with the size gt .5MB
21
First Attempt to Improve the Performance
HDF4 still performs 2 times better for the
hyperslabs gt 2MB. HDF5 shows nonlinear growth.
22
Current Behavior (HDF5 development branch)
HDF5 growth linear and performs about 10 times
better than HDF4.
23
Vdatas vs Compound Datasets
24
Vdatas and Compound Datasets

• In this test we created HDF4 files with Vdata and
  HDF5 files with compound dataset with sizes from
  1000 to 1000000 number of records
• float a short bfloat c3 char d
• write operation, write with packing data and
  partial read were tested.
• Test was performed on Linux platforms. We also
  looked into data conversion issues.

25
Writing Data (VSwrite and H5Dwrite)
Conversion does not affect HDF4 performance. It
does affect HDF5 ( more than in 15 times)
26
Writing Data (timing
includes packingVSpack and H5Tpack)
Data packing was added to the previous test. For
HDF5 we have very small effect.
27
Reading Two Fields
Unpacking slows down HDF4 significantly ( about 8
times) HDF5 was reading packed data in this test.
28
CERES Data File
29
Structure of CERES file
Vgroup CERES_ES8
Vgroup Geolocation Fields
Vgroup Data Fields
2
1
18
19
SDS
Vdata
SDS
Vdata
30
Ceres File

• Used H4toH5 converter to create an HDF5 version
  of the file
• 81MB (HDF4), 80MB (HDF5)
• 1 min 55 sec on Linux
• 3 min 56 sec on Solaris
• Benchmarks
• read up to 14 datasets (2148x660 floats)
• subsampling read two columns from the same
  datasets
• Benchmark was run on Solaris and Linux platforms

31
Reading CERES data on big and little - endian
machines
On Solaris platform, HDF5 was twice faster than
HDF4. On Linux (data conversion is on), HDF4 was
about 1.3-1.5 faster.
32
Subsetting CERES Data
Current version of HDF5 shows about 3 times
better performance.
33
Conclusion

• Goal tune HDF5 and give our users
  recommendations on its efficient usage
• Continue to study HDF4 and HDF5 performance
• try more platforms O2K, NT/Windows
• try other features (e.g. chunking, compression)
• specific HDF5 features (e.g. writing/reading big
  files, VL datatypes, compound datatypes,
  selections)
• Users input is necessary, send us access patterns
  you use!
• Results will be available _at_http//hdf.ncsa.uiuc.ed
  u