lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Michael McCandless (JIRA)" <>
Subject [jira] Commented: (LUCENE-1997) Explore performance of multi-PQ vs single-PQ sorting API
Date Fri, 30 Oct 2009 11:44:59 GMT


Michael McCandless commented on LUCENE-1997:

java version "1.6.0_14"
Java(TM) SE Runtime Environment (build 1.6.0_14-b08)
Java HotSpot(TM) 64-Bit Server VM (build 14.0-b16, mixed mode)

SunOS rhumba 5.11 snv_111b i86pc i386 i86pc Solaris

||Source||Seg size||Query||Tot hits||Sort||Top N||QPS old||QPS new||Pct change||
|wiki|log|1|1170209|rand int|10|27.08|27.52|{color:green}1.6%{color}|
|wiki|log|1|1170209|rand int|25|26.95|27.72|{color:green}2.9%{color}|
|wiki|log|1|1170209|rand int|50|27.54|27.03|{color:red}-1.9%{color}|
|wiki|log|1|1170209|rand int|100|27.08|24.19|{color:red}-10.7%{color}|
|wiki|log|1|1170209|rand int|500|26.13|26.88|{color:green}2.9%{color}|
|wiki|log|1|1170209|rand int|1000|24.30|25.91|{color:green}6.6%{color}|
|wiki|log|2|1088727|rand int|10|29.66|29.45|{color:red}-0.7%{color}|
|wiki|log|2|1088727|rand int|25|28.80|29.47|{color:green}2.3%{color}|
|wiki|log|2|1088727|rand int|50|29.57|30.21|{color:green}2.2%{color}|
|wiki|log|2|1088727|rand int|100|30.74|27.30|{color:red}-11.2%{color}|
|wiki|log|2|1088727|rand int|500|28.72|30.17|{color:green}5.0%{color}|
|wiki|log|2|1088727|rand int|1000|27.20|29.21|{color:green}7.4%{color}|
|wiki|log|<all>|5000000|rand int|10|33.43|31.09|{color:red}-7.0%{color}|
|wiki|log|<all>|5000000|rand int|25|33.24|31.77|{color:red}-4.4%{color}|
|wiki|log|<all>|5000000|rand int|50|32.89|31.68|{color:red}-3.7%{color}|
|wiki|log|<all>|5000000|rand int|100|32.04|21.33|{color:red}-33.4%{color}|
|wiki|log|<all>|5000000|rand int|500|31.25|30.59|{color:red}-2.1%{color}|
|wiki|log|<all>|5000000|rand int|1000|29.48|29.72|{color:green}0.8%{color}|
|random|log|<all>|5000000|rand int|10|32.61|31.64|{color:red}-3.0%{color}|
|random|log|<all>|5000000|rand int|25|32.98|31.79|{color:red}-3.6%{color}|
|random|log|<all>|5000000|rand int|50|32.63|30.89|{color:red}-5.3%{color}|
|random|log|<all>|5000000|rand int|100|32.08|21.60|{color:red}-32.7%{color}|
|random|log|<all>|5000000|rand int|500|30.48|30.65|{color:green}0.6%{color}|
|random|log|<all>|5000000|rand int|1000|28.95|29.44|{color:green}1.7%{color}|

This table compares trunk (= old) with the "inline int value directly into single PQ" approach
(=new).  So, a green result means the inline-single-PQ is faster; red means it's slower.

The results baffle me.  I would have expected for the 5M hits, with shallow topN, that the
diffs would be minor since the sub-leading cost should be in the noise for either approach,
and then as we go to fewer hits and deeper topN, that the inline-single-PQ approach would
be faster.  And, we still see strangeness at topN=100 where current trunk is always substantially
better.  The only difference between these ought to be the constant in front of the net number
of inserstions.  Strange!

> Explore performance of multi-PQ vs single-PQ sorting API
> --------------------------------------------------------
>                 Key: LUCENE-1997
>                 URL:
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Search
>    Affects Versions: 2.9
>            Reporter: Michael McCandless
>            Assignee: Michael McCandless
>         Attachments: LUCENE-1997.patch, LUCENE-1997.patch, LUCENE-1997.patch, LUCENE-1997.patch,
LUCENE-1997.patch, LUCENE-1997.patch, LUCENE-1997.patch, LUCENE-1997.patch, LUCENE-1997.patch
> Spinoff from recent "lucene 2.9 sorting algorithm" thread on java-dev,
> where a simpler (non-segment-based) comparator API is proposed that
> gathers results into multiple PQs (one per segment) and then merges
> them in the end.
> I started from John's multi-PQ code and worked it into
> contrib/benchmark so that we could run perf tests.  Then I generified
> the Python script I use for running search benchmarks (in
> contrib/benchmark/
> The script first creates indexes with 1M docs (based on
> SortableSingleDocSource, and based on wikipedia, if available).  Then
> it runs various combinations:
>   * Index with 20 balanced segments vs index with the "normal" log
>     segment size
>   * Queries with different numbers of hits (only for wikipedia index)
>   * Different top N
>   * Different sorts (by title, for wikipedia, and by random string,
>     random int, and country for the random index)
> For each test, 7 search rounds are run and the best QPS is kept.  The
> script runs singlePQ then multiPQ, and records the resulting best QPS
> for each and produces table (in Jira format) as output.

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message