[jira] Commented: (LUCENE-2722) Sep codec should store less in terms dict

["Michael McCandless (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/LUCENE-2722?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12924515#action_12924515 ] 

Michael McCandless commented on LUCENE-2722:
--------------------------------------------

Note that this is a change to the index format, but no core codecs actually use the [abstract] sep codec yet (LUCENE-1410 will change that).

So I ran a simple perf test using MockSep, which naively writes each vInt separately:

||Query||QPS trunk||QPS patch||Pct diff||||
|"unit state"|7.74|7.78|{color:green}0.4%{color}|
|uni*|17.12|17.31|{color:green}1.1%{color}|
|+unit +state|10.75|10.95|{color:green}1.9%{color}|
|unit*|29.71|30.38|{color:green}2.3%{color}|
|unit state|11.99|12.36|{color:green}3.1%{color}|
|un*d|65.00|67.85|{color:green}4.4%{color}|
|state|38.43|40.16|{color:green}4.5%{color}|
|u*d|19.92|21.94|{color:green}10.1%{color}|
|united~0.7|24.02|30.30|{color:green}26.2%{color}|
|united~0.6|5.26|7.00|{color:green}33.2%{color}|

It's a good speedup for the two fuzzy queries and also u*d...


> Sep codec should store less in terms dict
> -----------------------------------------
>
>                 Key: LUCENE-2722
>                 URL: https://issues.apache.org/jira/browse/LUCENE-2722
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Index
>            Reporter: Michael McCandless
>            Assignee: Michael McCandless
>             Fix For: 4.0
>
>         Attachments: LUCENE-2722.patch
>
>
> I'm working on improving Lucene's performance with int block codecs
> (FOR/PFOR), but in early perf testing I found that these codecs cause
> a big perf hit to those MTQs that need to scan many terms but don't
> end up accepting many of those terms (eg fuzzy, wildcard, regexp).
> This is because sep codec stores much more in the terms dict, since
> each file is separate, ie seek points for each of doc, frq, pos, pyl,
> skp files.
> So I'd like to shift these seek points to instead be stored in the doc
> file, except for the doc seek point itself.  Since a given query will
> always need to seek to the doc file, this does not add an extra seek.
> But it saves tons of vInt decodes for the next/seke intensive MTQs...

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