[
https://issues.apache.org/jira/browse/LUCENE-8673?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16760510#comment-16760510
]
Ignacio Vera commented on LUCENE-8673:
--------------------------------------
I have done changes recommended by [~jpountz] and I replaced the on-heap
selector from an IntroSelector to a radixSelector. These have provided another
nice bump on performance. Benchmarks looks like this now:
||Approach||Index time (sec): Dev||Index Time (sec): Base||Index Time:
Diff||Force merge time (sec): Dev||Force Merge time (sec): Base||Force Merge
Time: Diff||Index size (GB): Dev||Index size (GB): Base||Index Size:
Diff||Reader heap (MB): Dev||Reader heap (MB): Base||Reader heap: Diff||
|points|182.2s|227.9s|-20%|90.4s|143.1s|-37%|0.55|0.55| 0%|1.57|1.57| 0%|
|shapes|297.0s|624.4s|-52%|163.8s|549.3s|-70%|1.29|1.29| 0%|1.61|1.61| 0%|
|geo3d|210.3s|370.1s|-43%|104.3s|265.3s|-61%|0.75|0.75| 0%|1.58|1.58| 0%|
> Use radix partitioning when merging dimensional points
> ------------------------------------------------------
>
> Key: LUCENE-8673
> URL: https://issues.apache.org/jira/browse/LUCENE-8673
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Ignacio Vera
> Priority: Major
> Attachments: Geo3D.png, Geo3D.png, LatLonPoint.png, LatLonPoint.png,
> LatLonShape.png, LatLonShape.png
>
> Time Spent: 3h 50m
> Remaining Estimate: 0h
>
> Following the advise of [~jpountz] in LUCENE-8623I have investigated using
> radix selection when merging segments instead of sorting the data at the
> beginning. The results are pretty promising when running Lucene geo
> benchmarks:
>
> ||Approach||Index time (sec): Dev||Index Time (sec): Base||Index Time:
> Diff||Force merge time (sec): Dev||Force Merge time (sec): Base||Force Merge
> Time: Diff||Index size (GB): Dev||Index size (GB): Base||Index Size:
> Diff||Reader heap (MB): Dev||Reader heap (MB): Base||Reader heap: Diff
> |points|241.5s|235.0s| 3%|157.2s|157.9s|-0%|0.55|0.55| 0%|1.57|1.57| 0%|
> |shapes|416.1s|650.1s|-36%|306.1s|603.2s|-49%|1.29|1.29| 0%|1.61|1.61| 0%|
> |geo3d|261.0s|360.1s|-28%|170.2s|279.9s|-39%|0.75|0.75| 0%|1.58|1.58| 0%|
>
> edited: table formatting to be a jira table
>
> In 2D the index throughput is more or less equal but for higher dimensions
> the impact is quite big. In all cases the merging process requires much less
> disk space, I am attaching plots showing the different behaviour and I am
> opening a pull request.
>
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
