[
https://issues.apache.org/jira/browse/FLINK-18554?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17157258#comment-17157258
]
Truong Duc Kien commented on FLINK-18554:
-----------------------------------------
We investigated and came to the conclusion that it's almost impossible to
account for mmap's memory without CGroup: Linux will use as much memory as
allowed to mmap file. As a result, we've decided to just turn mmap_read off
when running on YARN without CGroup. Turning on mmap_read significantly lowers
the read latency when running Rocksb on ram disk, but our jobs can run fine
without it, so not a big loss.
Since mmap_read defaults to OFF, we also think it's fine for Flink to ignore
this edge case.
> Memory exceeds taskmanager.memory.process.size when enabling mmap_read for
> RocksDB
> ----------------------------------------------------------------------------------
>
> Key: FLINK-18554
> URL: https://issues.apache.org/jira/browse/FLINK-18554
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Configuration
> Affects Versions: 1.11.0
> Reporter: Truong Duc Kien
> Priority: Major
>
> We are testing Flink automatic memory management feature on Flink 1.11.
> However, YARN kept killing our containers due to the processes' physical
> memory exceeds the limit, although we have tuned the following configurations:
> {code:java}
> taskmanager.memory.process.size
> taskmanager.memory.managed.fraction
> {code}
> We suspect that it's because we have enabled mmap_read for RocksDB, since
> turning this options off seems to fix the issue. Maybe Flink automatic memory
> management is unable to account for the addition memory required when using
> mmap_read ?
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
