[
https://issues.apache.org/jira/browse/OPENNLP-1556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17843533#comment-17843533
]
ASF GitHub Bot commented on OPENNLP-1556:
-----------------------------------------
mawiesne opened a new pull request, #600:
URL: https://github.com/apache/opennlp/pull/600
Change
-
- adjusts TwoPassDataIndexer to make use of JDK's built-in
`CheckedOutputStream` / `CheckedInputStream` for checksum
([CRC32c](https://docs.oracle.com/en/java/javase/17/docs/api/java.base/java/util/zip/CRC32C.html))
computations
- removes untested class `HashSumEventStream` which is just a wrapper for
calling a slow toString() in Event to get some bytes to use for the computation
of a checksum
- provides a HashSumEventStream replacement: `ChecksumEventStream` which
makes use of the faster CRC32c checksum computation, avoiding cryptographic
hash functions such as MD5
- adds JUnit tests for ChecksumEventStream
Note(s)
-
1. A _full_ OpenNLP build is consistently about 3-4s faster on my local
machine (61s vs. 57s => ~ -7%) .
2. The effect should be greater when processing large(r) training corpora.
3. We should see a difference in a bare metal EVAL build run.
Tasks
-
Thank you for contributing to Apache OpenNLP.
In order to streamline the review of the contribution we ask you
to ensure the following steps have been taken:
### For all changes:
- [x] Is there a JIRA ticket associated with this PR? Is it referenced
in the commit message?
- [x] Does your PR title start with OPENNLP-XXXX where XXXX is the JIRA
number you are trying to resolve? Pay particular attention to the hyphen "-"
character.
- [x] Has your PR been rebased against the latest commit within the target
branch (typically main)?
- [x] Is your initial contribution a single, squashed commit?
### For code changes:
- [x] Have you ensured that the full suite of tests is executed via mvn
clean install at the root opennlp folder?
- [x] Have you written or updated unit tests to verify your changes?
- [ ] If adding new dependencies to the code, are these dependencies
licensed in a way that is compatible for inclusion under [ASF
2.0](http://www.apache.org/legal/resolved.html#category-a)?
- [ ] If applicable, have you updated the LICENSE file, including the main
LICENSE file in opennlp folder?
- [ ] If applicable, have you updated the NOTICE file, including the main
NOTICE file found in opennlp folder?
### For documentation related changes:
- [ ] Have you ensured that format looks appropriate for the output in which
it is rendered?
### Note:
Please ensure that once the PR is submitted, you check GitHub Actions for
build issues and submit an update to your PR as soon as possible.
> Improve speed of checksum computation in TwoPassDataIndexer
> -----------------------------------------------------------
>
> Key: OPENNLP-1556
> URL: https://issues.apache.org/jira/browse/OPENNLP-1556
> Project: OpenNLP
> Issue Type: Improvement
> Components: Machine Learning
> Affects Versions: 1.9.0, 2.0.0, 2.1.0, 2.2.0, 2.3.0, 2.3.1, 2.3.2, 2.3.3
> Reporter: Martin Wiesner
> Assignee: Martin Wiesner
> Priority: Major
> Fix For: 2.3.4
>
>
> For training ML models, all observations (Events) are indexed via
> {{TwoPassDataIndexer#index(ObjectStream<Event> eventStream)}}.
> When #index(..) is run, a tmp file is written and read in again. For the
> purpose of checksum validation, instances of HashSumEventStream are used to
> validate the content processed.
> Based on a rather slow toString() implementation in Event, a cryptographic
> (MD5) message digest is computed. This, however, is much slower than simply
> computing a checksum (such as a CRC32c value) for both directions
> (write/read). The (slowing) effect is more problematic when larger training
> corpora are (pre-)processed, that is, indexed in advance.
> Aims:
> - Speedup the (IO-bound) indexing part prior to the actual CPU-bound training
> phase.
> - Switch from MD5 to CRC32c, as there is *no* need for a cryptographic hash
> function here; it's simply a checksum that is required to decide whether all
> bytes written are the same bytes that are read.
> - Remove the untested class HashSumEventStream which is just a wrapper for
> calling a slow toString() in Event to get some bytes to use for the
> computation of a checksum / md.
> - Provide a replacement for HashSumEventStream, e.g. ChecksumEventStream that
> makes use of the faster CRC32c checksum computation, avoiding cryptographic
> hash functions such as MD5.
> - Make sure all existing tests hold.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
