Robert Muir created LUCENE-9403:
-----------------------------------
Summary: tune BufferedChecksum.DEFAULT_BUFFERSIZE
Key: LUCENE-9403
URL: https://issues.apache.org/jira/browse/LUCENE-9403
Project: Lucene - Core
Issue Type: Improvement
Reporter: Robert Muir
Attachments: LUCENE-9403.patch
This is currently set to 256 bytes, so that's the amount of data we pass to
crc.update() at once.
I tried different sizes with https://github.com/benalexau/hash-bench and JDK14:
{noformat}
HashBench.withArray crc32-jre 256 avgt 5 81.349 ± 8.364
ns/op
HashBench.withArray crc32-jre 512 avgt 5 95.204 ± 10.057
ns/op
HashBench.withArray crc32-jre 1024 avgt 5 120.081 ± 8.471
ns/op
HashBench.withArray crc32-jre 2048 avgt 5 173.505 ± 8.857
ns/op
HashBench.withArray crc32-jre 8192 avgt 5 487.721 ± 11.435
ns/op
{noformat}
based on this let's bump the buffersize from 256 to 1024? I think we want to
avoid huge buffers but still keep the CPU overhead low. It only impacts
ChecksumIndexInputs (e.g. speed of checkIntegrity() calls at merge) because
IndexOutputs do not need this buffer.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]