[jira] [Comment Edited] (TEZ-3349) Tez: PipelinedSorter - remove interface calls from the inner loop

["Gopal V (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/TEZ-3349?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15390245#comment-15390245 ] 

Gopal V edited comment on TEZ-3349 at 7/22/16 9:47 PM:
-------------------------------------------------------

The movabs + cmp is a type check - I'm guessing for a JIT inline cache, checking for the impl.

Doing that once per IndexedSortable::swap() seems to be excessive - using the SortSpan directly, could mean that op gets inlined into the QuickSort loops.


was (Author: gopalv):
The movabs + cmp is a type check - I'm guessing for a JIT inline cache, checking for the impl.

Doing that once per IndexedSortable::swap() seems to be excessive.

> Tez: PipelinedSorter - remove interface calls from the inner loop
> -----------------------------------------------------------------
>
>                 Key: TEZ-3349
>                 URL: https://issues.apache.org/jira/browse/TEZ-3349
>             Project: Apache Tez
>          Issue Type: Sub-task
>            Reporter: Gopal V
>         Attachments: movabs-check.png
>
>
> IndexedSortable is responsible for a huge degradation in sort performance in PipelinedSorter.
> With the TezRawComparator in place, the interface dispatch within indexed sortable is significantly affecting the speed of the sort impl.
> !movabs-check.png!



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)