[
https://issues.apache.org/jira/browse/FLINK-2653?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14802742#comment-14802742
]
ASF GitHub Bot commented on FLINK-2653:
---------------------------------------
Github user StephanEwen commented on the pull request:
https://github.com/apache/flink/pull/1115#issuecomment-141041498
You are right, in those drivers, it is handled in the wrong way. Probably
an artifact from the time before the `MultableObjectIterator` had both variants
of the `next()` method. It used to have only the reusing variant.
Clearly, this should be fixed.
Here is the guide that I try to follow when working with the mutable
objects:
- All `MultableObjectIterator`s have two variants of the `next()` method
- one for reuse, one without.
- The variant without reuse it crucial, as not every situation can work
with object reuse.
- The variant with reuse is optional, but should be implemented where
possible for performance.
- The task drivers (AllReduceDriver, ...) or algorithms (sorter / hasher)
know whether they want to work with reuse or not, and call the respective
method in that case.
> Enable object reuse in MergeIterator
> ------------------------------------
>
> Key: FLINK-2653
> URL: https://issues.apache.org/jira/browse/FLINK-2653
> Project: Flink
> Issue Type: Improvement
> Components: Distributed Runtime
> Affects Versions: master
> Reporter: Greg Hogan
>
> MergeIterator currently discards given reusable objects and simply returns a
> new object from the JVM heap. This inefficiency has a noticeable impact on
> garbage collection and runtime overhead (~5% overall performance by my
> measure).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
