lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Michael McCandless (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (LUCENE-5554) Add TermBulkScorer
Date Wed, 26 Mar 2014 10:30:17 GMT

    [ https://issues.apache.org/jira/browse/LUCENE-5554?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13947782#comment-13947782
] 

Michael McCandless commented on LUCENE-5554:
--------------------------------------------

I think both approaches amount to hotspot voodoo.

But I agree Rob's patch is less voodoo-ey, and I ran a perf test where base = just before
LUCENE-5487 was committed, and comp = just after it was committed PLUS Rob's patch:

{noformat}
                    Task    QPS base      StdDev    QPS comp      StdDev                Pct
diff
              AndHighLow      300.56      (3.0%)      291.94      (2.4%)   -2.9% (  -8% -
   2%)
                  Fuzzy1       43.83      (2.9%)       42.85      (4.0%)   -2.2% (  -8% -
   4%)
                  Fuzzy2       34.41      (2.5%)       33.65      (3.5%)   -2.2% (  -8% -
   3%)
               OrHighLow       17.16      (1.7%)       16.91      (2.8%)   -1.4% (  -5% -
   3%)
            OrHighNotLow       22.29      (1.8%)       22.02      (2.8%)   -1.2% (  -5% -
   3%)
           OrHighNotHigh       10.41      (1.7%)       10.29      (2.6%)   -1.2% (  -5% -
   3%)
           OrNotHighHigh        8.40      (1.9%)        8.31      (2.7%)   -1.2% (  -5% -
   3%)
               OrHighMed       23.29      (1.7%)       23.03      (2.6%)   -1.1% (  -5% -
   3%)
             AndHighHigh       23.08      (2.2%)       22.85      (2.0%)   -1.0% (  -5% -
   3%)
            OrHighNotMed       26.20      (1.5%)       25.94      (2.6%)   -1.0% (  -4% -
   3%)
                 Respell       37.25      (3.0%)       36.90      (3.7%)   -1.0% (  -7% -
   5%)
              AndHighMed       27.81      (2.5%)       27.56      (2.0%)   -0.9% (  -5% -
   3%)
            OrNotHighLow       19.65      (1.4%)       19.54      (2.6%)   -0.6% (  -4% -
   3%)
        HighSloppyPhrase        2.67      (6.8%)        2.66      (7.2%)   -0.5% ( -13% -
  14%)
            OrNotHighMed       18.70      (1.5%)       18.61      (2.5%)   -0.5% (  -4% -
   3%)
         MedSloppyPhrase        2.87      (5.0%)        2.86      (4.3%)   -0.4% (  -9% -
   9%)
              OrHighHigh        7.55      (2.0%)        7.53      (3.2%)   -0.3% (  -5% -
   5%)
                Wildcard       15.12      (3.8%)       15.08      (3.6%)   -0.3% (  -7% -
   7%)
              HighPhrase        3.25      (7.6%)        3.26      (6.6%)    0.5% ( -12% -
  15%)
         LowSloppyPhrase       35.48      (2.3%)       35.66      (2.4%)    0.5% (  -4% -
   5%)
             MedSpanNear       24.07      (4.0%)       24.26      (3.6%)    0.8% (  -6% -
   8%)
               LowPhrase       10.25      (3.2%)       10.34      (2.3%)    0.9% (  -4% -
   6%)
                 Prefix3       72.06      (2.7%)       72.71      (3.2%)    0.9% (  -4% -
   6%)
                  IntNRQ        2.79      (5.0%)        2.82      (5.3%)    1.1% (  -8% -
  11%)
             LowSpanNear        8.19      (4.0%)        8.29      (4.2%)    1.2% (  -6% -
   9%)
            HighSpanNear        6.15      (5.8%)        6.23      (5.3%)    1.3% (  -9% -
  13%)
               MedPhrase      157.17      (6.9%)      159.52      (6.9%)    1.5% ( -11% -
  16%)
                 MedTerm       70.27     (11.4%)       71.81     (12.4%)    2.2% ( -19% -
  29%)
                HighTerm       49.45     (12.8%)       50.94     (14.9%)    3.0% ( -21% -
  35%)
                 LowTerm      209.81     (13.9%)      218.70     (11.2%)    4.2% ( -18% -
  33%)
{noformat}

Ie, we want everything to be 0%, and so it looks like Rob's patch makes hotspot happy again
and is more general than my first patch.

So +1 to commit.

> Add TermBulkScorer
> ------------------
>
>                 Key: LUCENE-5554
>                 URL: https://issues.apache.org/jira/browse/LUCENE-5554
>             Project: Lucene - Core
>          Issue Type: Improvement
>          Components: core/search
>            Reporter: Michael McCandless
>            Assignee: Michael McCandless
>             Fix For: 4.8, 5.0
>
>         Attachments: LUCENE-5554.patch, LUCENE-5554.patch
>
>
> Hotspot was unhappy with the changes in LUCENE-5487, e.g.: http://people.apache.org/~mikemccand/lucenebench/OrHighHigh.html
> But it looks like we can get the performance back by making a dedicated BulkScorer for
TermQuery.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org


Mime
View raw message