[
https://issues.apache.org/jira/browse/TIKA-5055?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18139997#comment-18139997
]
Henry Lindeman commented on TIKA-5055:
--------------------------------------
Ok all I have that isn't on 5057 is the shift-based BitReader (which was a
significant speedup) and the GUID-UUID cache (which was less significant).
Rebased those changes on your branch and measured w/ your perf test
||Build||MB per parse||ms per parse||Load: wall||GC pauses||Full GCs||
|Tim|6.00|3.06|0.62 s|178|9|
|+ shift reads|5.03 (−16.2%)|2.79 (−8.5%)|0.54 s (−12.3%)|80 (−55%)|0|
|+ GUID cache|4.85 (−3.7%)|2.73 (−2.2%)|0.50 s (−7.2%)|76 (−5%)|0|
lmk what I should do gitwise
> Improve onenote performance
> ---------------------------
>
> Key: TIKA-5055
> URL: https://issues.apache.org/jira/browse/TIKA-5055
> Project: Tika
> Issue Type: Improvement
> Components: parser
> Reporter: Henry Lindeman
> Priority: Minor
>
> We rolled out the onenote stuff to a customer, and performance degraded a
> bunch from GC pauses.
> ||Sample date (2026)||p95 GC pause||p99 GC pause||
> ||Oct 8|0.49|0.62|
> ||Oct 13|0.33|0.46|
> ||Oct 15|0.35|0.5|
> ||Oct 16|0.4|0.66|
> ||Oct 19|30.85|66.48|
> ||Oct 22|34.21|65.23|
> ||Nov 5|31.73|109.47|
> I haven't gotten access to the customer documents to test on but measuring
> the test docs I've been working with had it making over 500 gc calls in 1000
> parses of the section-2 document, which contains a large image. Anyway, it
> looks like storing an image as List<Byte> is real bad, compared to byte[]. PR
> incoming
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
