[
https://issues.apache.org/jira/browse/HIVE-15573?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15816767#comment-15816767
]
Gopal V edited comment on HIVE-15573 at 1/11/17 1:31 AM:
---------------------------------------------------------
The timings for Map 1 went from 85s -> 13s, when vectorization (incorrectly
bucketed) applied to this ReduceSink.
Some part of that gain will be negated by the proper bucketing, with an
estimated ~2x speedup to the mapper tasks when this is fixed.
was (Author: gopalv):
The timings for Map 1 went from 85s -> 13s, when vectorization (incorrectly
bucketed) applied to this ReduceSink.
> Vectorization: ACID shuffle ReduceSink is not specialized
> ----------------------------------------------------------
>
> Key: HIVE-15573
> URL: https://issues.apache.org/jira/browse/HIVE-15573
> Project: Hive
> Issue Type: Improvement
> Components: Transactions, Vectorization
> Affects Versions: 2.2.0
> Reporter: Gopal V
> Attachments: screenshot-1.png
>
>
> The ACID shuffle disabled murmur hash for the shuffle, due to the bucketing
> requirements demanding the writable hashcode for the shuffles.
> {code}
> boolean useUniformHash = desc.getReducerTraits().contains(UNIFORM);
> if (!useUniformHash) {
> return false;
> }
> {code}
> This check protects the fast ReduceSink ops from being used in ACID inserts.
> A specialized case for the following pattern will make ACID insert much
> faster.
> {code}
> Reduce Output Operator
> sort order:
> Map-reduce partition columns: _col0 (type: bigint)
> value expressions: ....
> {code}
> !screenshot-1.png!
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
