[
https://issues.apache.org/jira/browse/FLINK-10995?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16942129#comment-16942129
]
Piotr Nowojski commented on FLINK-10995:
----------------------------------------
After modifying the benchmarks to actually use this code path performance
improvement is easily visible:
http://codespeed.dak8s.net:8000/timeline/#/?exe=1,3&ben=networkBroadcastThroughput&env=2&revs=200&equid=off&quarts=on&extr=on
> Copy intermediate serialization results only once for broadcast mode
> --------------------------------------------------------------------
>
> Key: FLINK-10995
> URL: https://issues.apache.org/jira/browse/FLINK-10995
> Project: Flink
> Issue Type: Sub-task
> Components: Runtime / Task
> Affects Versions: 1.8.0
> Reporter: zhijiang
> Assignee: zhijiang
> Priority: Minor
> Labels: pull-request-available
> Time Spent: 40m
> Remaining Estimate: 0h
>
> The emitted records from operator would be firstly serialized into
> intermediate bytes array in {{RecordSerializer}}, then copy the intermediate
> results into target buffers for different sub partitions. For broadcast
> mode, the same intermediate results would be copied as many times as the
> number of sub partitions, and this would affect the performance seriously in
> large scale jobs.
> We can copy to only one target buffer which would be shared by all the sub
> partitions to reduce the overheads. For emitting latency marker in broadcast
> mode, we should flush the previous shared target buffers first, and then
> request a new buffer for the target sub partition to send latency marker.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
