[
https://issues.apache.org/jira/browse/SPARK-1946?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Zhihui updated SPARK-1946:
--------------------------
Description:
Because creating TaskSetManager and registering executors are asynchronous, if
running job without enough executors, it will lead to some issues
* early stages' tasks run without preferred locality.
* the default parallelism in yarn is based on number of executors,
* the number of intermediate files per node for shuffle (this can bring the
node down btw)
* and amount of memory consumed on a node for rdd MEMORY persisted data (making
the job fail if disk is not specified : like some of the mllib algos ?)
* and so on ...
(thanks [~mridulm80] 's [comments |
https://github.com/apache/spark/pull/900#issuecomment-45780405])
A simple solution is sleeping few seconds in application, so that executors
have enough time to register.
A better way is to make DAGScheduler submit stage after a few of executors have
been registered by configuration properties.
\# submit stage only after successfully registered executors arrived the ratio,
default value 0 in Standalone mode and 0.9 in Yarn mode
spark.scheduler.minRegisteredRatio = 0.8
\# whatever registered number is arrived, submit stage after the
maxRegisteredWaitingTime(millisecond), default value 10000
spark.scheduler.maxRegisteredWaitingTime = 5000
was:
Because creating TaskSetManager and registering executors are asynchronous, if
running job without enough executors, it will lead to some issues
* early stages' tasks run without preferred locality.
* the default parallelism in yarn is based on number of executors,
* the number of intermediate files per node for shuffle (this can bring the
node down btw)
* and amount of memory consumed on a node for rdd MEMORY persisted data (making
the job fail if disk is not specified : like some of the mllib algos ?)
* and so on ...
(thanks [~mridulm80] 's [comments |
https://github.com/apache/spark/pull/900#issuecomment-45780405])
A simple solution is sleeping few seconds in application, so that executors
have enough time to register.
A better way is to make DAGScheduler submit stage after a few of executors have
been registered by configuration properties.
\# submit stage only after successfully registered executors arrived the
number, default value 0
spark.executor.minRegisteredNum = 20
\# whatever registeredRatio is arrived, submit stage after the
maxRegisteredWaitingTime(millisecond), default value 10000
spark.executor.maxRegisteredWaitingTime = 5000
> Submit stage after executors have been registered
> -------------------------------------------------
>
> Key: SPARK-1946
> URL: https://issues.apache.org/jira/browse/SPARK-1946
> Project: Spark
> Issue Type: Improvement
> Components: Spark Core
> Affects Versions: 1.0.0
> Reporter: Zhihui
> Attachments: Spark Task Scheduler Optimization Proposal.pptx
>
>
> Because creating TaskSetManager and registering executors are asynchronous,
> if running job without enough executors, it will lead to some issues
> * early stages' tasks run without preferred locality.
> * the default parallelism in yarn is based on number of executors,
> * the number of intermediate files per node for shuffle (this can bring the
> node down btw)
> * and amount of memory consumed on a node for rdd MEMORY persisted data
> (making the job fail if disk is not specified : like some of the mllib algos
> ?)
> * and so on ...
> (thanks [~mridulm80] 's [comments |
> https://github.com/apache/spark/pull/900#issuecomment-45780405])
> A simple solution is sleeping few seconds in application, so that executors
> have enough time to register.
> A better way is to make DAGScheduler submit stage after a few of executors
> have been registered by configuration properties.
> \# submit stage only after successfully registered executors arrived the
> ratio, default value 0 in Standalone mode and 0.9 in Yarn mode
> spark.scheduler.minRegisteredRatio = 0.8
> \# whatever registered number is arrived, submit stage after the
> maxRegisteredWaitingTime(millisecond), default value 10000
> spark.scheduler.maxRegisteredWaitingTime = 5000
--
This message was sent by Atlassian JIRA
(v6.2#6252)
