[
https://issues.apache.org/jira/browse/LUCENE-4620?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13552573#comment-13552573
]
Michael McCandless commented on LUCENE-4620:
--------------------------------------------
This change seemed to lose a bit of performance: look at 1/11/2013 on
http://people.apache.org/~mikemccand/lucenebench/TermDateFacets.html
But, that tests just one dimension (Date), with only 3 ords per doc,
so I had assumed that this just wasn't enough ints being decoded to
see the gains from this bulk decoding.
So, I modified luceneutil to have more facets per doc (avg ~25 ords
per doc across 9 dimensions; 2.5M unique ords), and the results are
still slower:
{noformat}
Task QPS base StdDev QPS comp StdDev Pct
diff
HighTerm 3.62 (2.5%) 3.24 (1.0%) -10.5% ( -13% -
-7%)
MedTerm 7.34 (1.7%) 6.78 (0.9%) -7.6% ( -10% -
-5%)
LowTerm 14.92 (1.6%) 14.32 (1.2%) -4.0% ( -6% -
-1%)
PKLookup 181.47 (4.7%) 183.04 (5.3%) 0.9% ( -8% -
11%)
{noformat}
This is baffling ... not sure what's up. I would expect some gains
given that the micro-benchmark showed sizable decode improvements. It
must somehow be that decode cost is a minor part of facet counting?
(which is not a good sign!: it should be a big part of it...)
> Explore IntEncoder/Decoder bulk API
> -----------------------------------
>
> Key: LUCENE-4620
> URL: https://issues.apache.org/jira/browse/LUCENE-4620
> Project: Lucene - Core
> Issue Type: Improvement
> Components: modules/facet
> Reporter: Shai Erera
> Assignee: Shai Erera
> Fix For: 4.1, 5.0
>
> Attachments: LUCENE-4620.patch, LUCENE-4620.patch, LUCENE-4620.patch
>
>
> Today, IntEncoder/Decoder offer a streaming API, where you can encode(int)
> and decode(int). Originally, we believed that this layer can be useful for
> other scenarios, but in practice it's used only for writing/reading the
> category ordinals from payload/DV.
> Therefore, Mike and I would like to explore a bulk API, something like
> encode(IntsRef, BytesRef) and decode(BytesRef, IntsRef). Perhaps the Encoder
> can still be streaming (as we don't know in advance how many ints will be
> written), dunno. Will figure this out as we go.
> One thing to check is whether the bulk API can work w/ e.g. facet
> associations, which can write arbitrary byte[], and so may decoding to an
> IntsRef won't make sense. This too we'll figure out as we go. I don't rule
> out that associations will use a different bulk API.
> At the end of the day, the requirement is for someone to be able to configure
> how ordinals are written (i.e. different encoding schemes: VInt, PackedInts
> etc.) and later read, with as little overhead as possible.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
