[
https://issues.apache.org/jira/browse/SPARK-41449?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Shay Elbaz updated SPARK-41449:
-------------------------------
Description:
Since the total/max number of executor is constant throughout the application -
in dynamic or static allocation - there is loose control over how much GPUs
will be requested from the resource manager.
For example, if an application needs 500 executors for the ETL part (with N
cores each), but it needs - *or allowed -* only 50 GPUs for the DL part, in
practice it will request at least 500 GPUs from the RM, since
`spark.executor.instances` is set to 500. This leads to resource management
challenges in multi tenant environments.
A quick workaround is to repartition the RDD to 50 partitions just before
switching resources, but it has obvious downsides.
It would be very helpful if the total/max number of executors could also be
configured in the Resource Profile.
was:
Since the total/max number of executor is constant throughout the application -
in dynamic or static allocation - there is loose control over how much GPUs
will be requested from the resource manager.
For example, if an application needs 500 executors for the ETL part (with N
cores each), but it needs - *or allowed -* only 50 GPUs for the DL part, in
practice it will request at least 500 GPUs from the RM, since
`spark.executor.instances` is set to 500. This leads to resource management
challenges in multi tenant environments.
A quick workaround is to repartition the RDD to 50 partitions just before
switching resources, but it has obvious downsides.
It would be very helpful if the total/max number of executors could be also
configured in the Resource Profile.
> Stage level scheduling, allow to change number of executors
> -----------------------------------------------------------
>
> Key: SPARK-41449
> URL: https://issues.apache.org/jira/browse/SPARK-41449
> Project: Spark
> Issue Type: Improvement
> Components: Scheduler
> Affects Versions: 3.3.0, 3.3.1
> Reporter: Shay Elbaz
> Priority: Major
> Labels: scheduler
>
> Since the total/max number of executor is constant throughout the application
> - in dynamic or static allocation - there is loose control over how much GPUs
> will be requested from the resource manager.
> For example, if an application needs 500 executors for the ETL part (with N
> cores each), but it needs - *or allowed -* only 50 GPUs for the DL part, in
> practice it will request at least 500 GPUs from the RM, since
> `spark.executor.instances` is set to 500. This leads to resource management
> challenges in multi tenant environments.
> A quick workaround is to repartition the RDD to 50 partitions just before
> switching resources, but it has obvious downsides.
> It would be very helpful if the total/max number of executors could also be
> configured in the Resource Profile.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
