[
https://issues.apache.org/jira/browse/SPARK-4922?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14268826#comment-14268826
]
Jongyoul Lee commented on SPARK-4922:
-------------------------------------
[~andrewor14] Hi, I have a basic question about your idea. I'm using
fine-grained mesos for running my jobs. that mode already allocate resources
dynamically when task scheduler wants. What you think the difference is between
your idea and fine-grained mode? Unlike coarse-grained mode, fine-grained mode
adjusts # of cores for a executor and enables to make two more executor on each
slave. I think if we set # of cores for each mesos executor in a configuration
on fine-grained mode - now, only one core fixed for each executor -, we can
satisfy dynamic allocation idea. and I read SPARK-4751, and I'll handle this
issue via using fine-grain mode. And how do you think you adjust resources? new
API for increasing or decreasing cores or just use {{spark.cores.max}}?
> Support dynamic allocation for coarse-grained Mesos
> ---------------------------------------------------
>
> Key: SPARK-4922
> URL: https://issues.apache.org/jira/browse/SPARK-4922
> Project: Spark
> Issue Type: Bug
> Components: Mesos
> Affects Versions: 1.2.0
> Reporter: Andrew Or
> Priority: Critical
>
> This brings SPARK-3174, which provided dynamic allocation of cluster
> resources to Spark on YARN applications, to Mesos coarse-grained mode.
> Note that the translation is not as trivial as adding a code path that
> exposes the request and kill mechanisms as we did for YARN is SPARK-3822.
> This is because Mesos coarse-grained mode schedules on the notion of setting
> the number of cores allowed for an application (as in standalone mode)
> instead of number of executors (as in YARN mode). For more detail, please see
> SPARK-4751.
> If you intend to work on this, please provide a detailed design doc!
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
