HDF5: Empiric for Optimal Chunk Size #916

ax3l · 2021-01-29T23:19:12Z

This ports a prior empirical algorithm from libSplash to determine an optimal (large) chunk size for an HDF5 dataset based on its datatype and global extent.

Original implementation by Felix Schmitt @f-schmitt (ZIH, TU Dresden) in libSplash.

Original source:

Close #406
Related to #898 (improve HDF5 baseline performance)
Required for #510: basis to extend resizable data sets (#829) to HDF5

To Do

pass tests
measure performance (h5chunks_cori.sbatch.txt)
- Cori seems to be still slow, also when playing further with OPENPMD_HDF5_INDEPENDENT="OFF" and OPENPMD_HDF5_ALIGNMENT="1048576" for our 8_benchmark_parallel -w benchmark
measure filesize overhead

Sample & bin directory du -hs bin samples with:

`OPENPMD_HDF5_CHUNKS`	size
"auto"/unset	13M + 1.6G
"none"	11M + 1.6G

on my laptop (4KiB blocksize).

With "auto" the MPI.8_benchmark_parallel test is significantly slower. Changing the 4D test to 3D brings the difference down to about 20% slowdown #1010. (Maybe the 4th, 10-element dimension is sub-ideal for chunking?)

Follow-Ups

allow definition on a per-dataset basis
allow passing user-defined per-dataset Byte arrays for chunks

ax3l · 2021-01-31T21:22:14Z

At the moment, a naïve port of the logic causes a deadlock in the MPI tests.

Stuck only in hipace_like_write with

dims: 300 20
optimal chunks found: 256 16

After some tests, it looks like adding chunking makes the dataset declaration H5Dcreate collective in that case, urgh... (It's actually declared as always collective.)

Notes:

Chunk size cannot exceed the size of a fixed-size dataset. For example, a dataset consisting of a 5x4 fixed-size array cannot be defined with 10x10 chunks.

ax3l · 2021-01-31T21:56:55Z

src/IO/HDF5/HDF5IOHandler.cpp

+            //for( auto const& val : parameters.chunkSize )
+            //    chunk_dims.push_back(static_cast< hsize_t >(val));
+
+            herr_t status = H5Pset_chunk(datasetCreationProperty, chunk_dims.size(), chunk_dims.data());


Fun fact: H5Pset_chunk_opts (HDF5 1.10.0+):

H5Pset_chunk_opts is used to specify storage options for chunks on the edge of a dataset’s dataspace. This capability allows the user to tune performance in cases where the dataset size may not be a multiple of the chunk size and the handling of partial edge chunks can impact performance.

franzpoeschel · 2021-02-18T12:58:21Z

I've added some scaffolding for JSON options in HDF5

ax3l · 2021-06-10T05:05:53Z

Finished the global options (JSON & env) to disable chucking when needed (mainly for HiPACE's legacy pipeline + potential regressions).

ax3l · 2021-06-14T00:09:48Z

It's a bit concerning that the clang sanitizer run parallel benchmark (8) runs into a time-out with the new patch:

15/32 Test #15: MPI.8_benchmark_parallel ...............***Timeout 1500.17 sec

But I don't see an immediately relation. 'll turn the chunking off for this one.

src/IO/HDF5/HDF5IOHandler.cpp

franzpoeschel · 2021-06-18T13:33:37Z

I tried out whether this PR alone already enables extensible datasets in HDF5, apparently not:

HDF5-DIAG: Error detected in HDF5 (1.10.4) thread 0:
  #000: ../../../src/H5D.c line 906 in H5Dset_extent(): unable to set extend dataset
    major: Dataset
    minor: Unable to initialize object 
  #001: ../../../src/H5Dint.c line 2615 in H5D__set_extent(): unable to modify size of dataspace
    major: Dataset
    minor: Unable to initialize object
  #002: ../../../src/H5S.c line 2005 in H5S_set_extent(): dimension cannot exceed the existing maximal size (new: 10 max: 5)
    major: Dataspace
    minor: Bad value

According to the documentation, only certain kinds of datasets can have their extents extended:

A chunked dataset with unlimited dimensions
A chunked dataset with fixed dimensions if the new dimension sizes are less than the maximum sizes set with maxdims (see H5Screate_simple)

Our datasets seem to fall into the second category. I guess, for extensible datasets, we would have to create unlimited datasets from the beginning?

ax3l · 2021-06-21T22:02:59Z

In order to make a dataset resizable, we need to pass H5F_UNLIMITED into the dimensions of H5Screate_simple that should be resizable. Yes, we need chunking (this PR) to be able to use this (ref / PR).

Setting H5F_UNLIMITED on dimensions can have a negative performance implication, so users need to declare that intent upfront and then we just make all dims of a data set resizable.

@f-schmitt

This ports a prior empirical algorithm from libSplash to determine an optimal (large) chunk size for an HDF5 dataset based on its datatype and global extent. Original implementation by Felix Schmitt @f-schmitt (ZIH, TU Dresden) in [libSplash](https://github.com/ComputationalRadiationPhysics/libSplash). Original source: - https://github.com/ComputationalRadiationPhysics/libSplash/blob/v1.7.0/src/DCDataSet.cpp - https://github.com/ComputationalRadiationPhysics/libSplash/blob/v1.7.0/src/include/splash/core/DCHelper.hpp Co-authored-by: Felix Schmitt <[email protected]>

The parallel, independent I/O pattern here is corner-case for what HDF5 can support, due to non-collective declarations of data sets. Testing shows that it does not work with chunking.

Runs into timeout for unclear reasons with this patch: ``` 15/32 Test openPMD#15: MPI.8_benchmark_parallel ...............***Timeout 1500.17 sec ```

Co-authored-by: Franz Pöschel <[email protected]>

ax3l · 2022-01-14T23:19:34Z

docs/source/details/backendconfig.rst

+
+A full configuration of the HDF5 backend:
+
+.. literalinclude:: hdf5.json


@franzpoeschel I just realized I forgot to add a hdf5.json file here 🤪

Ah, I accidentally named it json.json

ax3l · 2022-01-14T23:20:12Z

docs/source/details/backendconfig.rst

+All keys found under ``hdf5.dataset`` are applicable globally (future: as well as per dataset).
+Explanation of the single keys:
+
+* ``adios2.dataset.chunks``: This key contains options for data chunking via `H5Pset_chunk <https://support.hdfgroup.org/HDF5/doc/RM/H5P/H5Pset_chunk.htm>`__.


Ouch, that should read hdf5.dataset.chunks...

ax3l added the backend: HDF5 label Jan 29, 2021

ax3l requested a review from guj January 29, 2021 23:21

ax3l force-pushed the topic-hdf5OptChunkSize branch from 287d7ee to fdbbf4f Compare January 29, 2021 23:23

ax3l commented Jan 31, 2021

View reviewed changes

ax3l force-pushed the topic-hdf5OptChunkSize branch 2 times, most recently from a625114 to 83695d8 Compare February 1, 2021 07:43

ax3l requested a review from franzpoeschel February 1, 2021 07:43

ax3l mentioned this pull request Feb 2, 2021

HDF5 Auxiliary: .cpp file #917

Merged

ax3l force-pushed the topic-hdf5OptChunkSize branch from 83695d8 to 04c0164 Compare February 2, 2021 20:15

ax3l force-pushed the topic-hdf5OptChunkSize branch from 04c0164 to 44f72a4 Compare February 16, 2021 19:09

ax3l force-pushed the topic-hdf5OptChunkSize branch from 0c2a55a to 34cde08 Compare April 6, 2021 06:03

ax3l force-pushed the topic-hdf5OptChunkSize branch from 34cde08 to 22f70f0 Compare June 10, 2021 05:03

ax3l changed the title ~~[WIP] HDF5: Empiric for Optimal Chunk Size~~ HDF5: Empiric for Optimal Chunk Size Jun 10, 2021

ax3l force-pushed the topic-hdf5OptChunkSize branch 4 times, most recently from 368c2c4 to aea7e96 Compare June 11, 2021 19:35

ax3l mentioned this pull request Jun 11, 2021

[WIP] Parallel HDF5: 4MB Alignment & Buffer #898

Open

ax3l force-pushed the topic-hdf5OptChunkSize branch from aea7e96 to 2d34cf6 Compare June 12, 2021 20:39

ax3l commented Jun 14, 2021

View reviewed changes

src/IO/HDF5/HDF5IOHandler.cpp Show resolved Hide resolved

ax3l commented Jun 14, 2021

View reviewed changes

src/IO/HDF5/HDF5IOHandler.cpp Outdated Show resolved Hide resolved

ax3l force-pushed the topic-hdf5OptChunkSize branch from d834f9f to ff1d13d Compare June 14, 2021 23:06

franzpoeschel requested changes Jun 18, 2021

View reviewed changes

src/IO/HDF5/HDF5IOHandler.cpp Outdated Show resolved Hide resolved

src/IO/HDF5/HDF5IOHandler.cpp Outdated Show resolved Hide resolved

franzpoeschel and others added 4 commits June 23, 2021 14:19

Add scaffolding for JSON options in HDF5

2e01d59

HDF5: Finish Chunking JSON/Env control

5610b2d

HiPACE (legacy) pipeline: no chunking

f1adbfc

The parallel, independent I/O pattern here is corner-case for what HDF5 can support, due to non-collective declarations of data sets. Testing shows that it does not work with chunking.

CI: no HDF5 Chunking with Sanitizer

16c14c7

Runs into timeout for unclear reasons with this patch: ``` 15/32 Test openPMD#15: MPI.8_benchmark_parallel ...............***Timeout 1500.17 sec ```

ax3l force-pushed the topic-hdf5OptChunkSize branch from ff1d13d to 16c14c7 Compare June 23, 2021 21:20

Apply suggestions from code review

d49f160

Co-authored-by: Franz Pöschel <[email protected]>

ax3l merged commit 8c2d9ce into openPMD:dev Jun 24, 2021

ax3l deleted the topic-hdf5OptChunkSize branch June 24, 2021 02:31

ax3l mentioned this pull request Jun 25, 2021

HDF5: Resizable Datasets #1020

Merged

ax3l mentioned this pull request Jul 8, 2021

Docs: makeConstant & Parallel HDF5 #1041

Merged

ax3l commented Jan 14, 2022

View reviewed changes


		A full configuration of the HDF5 backend:

		.. literalinclude:: hdf5.json

HDF5: Empiric for Optimal Chunk Size #916

HDF5: Empiric for Optimal Chunk Size #916

Uh oh!

Conversation

ax3l commented Jan 29, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

To Do

Follow-Ups

Uh oh!

ax3l commented Jan 31, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ax3l Jan 31, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

franzpoeschel commented Feb 18, 2021

Uh oh!

ax3l commented Jun 10, 2021

Uh oh!

ax3l commented Jun 14, 2021

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

franzpoeschel commented Jun 18, 2021

Uh oh!

ax3l commented Jun 21, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ax3l Jan 14, 2022

Choose a reason for hiding this comment

Uh oh!

ax3l Jan 14, 2022

Choose a reason for hiding this comment

Uh oh!

ax3l Jan 14, 2022

Choose a reason for hiding this comment

Uh oh!

ax3l Jan 14, 2022

Choose a reason for hiding this comment

Uh oh!

ax3l Jan 14, 2022

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

ax3l commented Jan 29, 2021 •

edited

Loading

ax3l commented Jan 31, 2021 •

edited

Loading

ax3l Jan 31, 2021 •

edited

Loading

ax3l commented Jun 21, 2021 •

edited

Loading