Github user tgravescs commented on the issue:
https://github.com/apache/spark/pull/18388
Haven't looked at the path in detail yet. High level questions/thoughts.
So you say the memory usage is by the netty chunks, so my assumption is
this is during the actual transfer? failing the open blocks isn't necessarily
going to solve that. If a bunch of reducers all due open blocks at once, it
won't reject any and when they all start to transfer it could still run out of
memory. It could help in the normal case where some run openblocks while
other transfers going on though. Have you been running this patch, what are
results?
So an alternative to this is limiting the number of blocks each reducer is
fetching at once. Instead of calling open blocks with 500 at once, do them in
chunks of say 20. We are working on a patch for that and should have it
available in the next couple days. This again though doesn't guarantee it but
it allows you to throttle down the # of blocks each reducer would get at once.
MapReduce/Tez actually do this with a lot of success.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
