[
https://issues.apache.org/jira/browse/SPARK-19090?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15813689#comment-15813689
]
nirav patel edited comment on SPARK-19090 at 1/10/17 3:14 AM:
--------------------------------------------------------------
[~jerryshao]
"spark.executor.cores" is to tell spark AM to request no of vcores from Yarn
per container. I think spark AM makes correct decision when dynamic allocation
is off but when its on it ignores spark.executor.cores value. I think DRF has
nothing to do with this issue. Following are AM logs from two different runs.
Run 1:
spark.dynamicAllocation.enabled = true
spark.executor.instances = 6
spark.executor.cores = 5
Dynamic allocation = true
17/01/09 19:05:49 INFO yarn.YarnRMClient: Registering the ApplicationMaster
17/01/09 19:05:49 INFO yarn.YarnAllocator: Will request 6 executor containers,
each with 1 cores and 11000 MB memory including 1000 MB overhead
17/01/09 19:05:49 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:1, disks:0.0>)
17/01/09 19:05:49 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:1, disks:0.0>)
17/01/09 19:05:49 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:1, disks:0.0>)
17/01/09 19:05:49 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:1, disks:0.0>)
17/01/09 19:05:49 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:1, disks:0.0>)
17/01/09 19:05:49 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:1, disks:0.0>)
Run 2:
spark.dynamicAllocation.enabled = false
spark.executor.instances = 6
spark.executor.cores = 5
17/01/09 19:01:39 INFO yarn.YarnAllocator: Will request 6 executor containers,
each with 5 cores and 11000 MB memory including 1000 MB overhead
17/01/09 19:01:39 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:5, disks:0.0>)
17/01/09 19:01:39 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:5, disks:0.0>)
17/01/09 19:01:39 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:5, disks:0.0>)
17/01/09 19:01:39 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:5, disks:0.0>)
17/01/09 19:01:39 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:5, disks:0.0>)
17/01/09 19:01:39 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:5, disks:0.0>)
I can verify same fact via Spark UI when job is running that with dynamic
allocation there is only 1 task running per executor.
was (Author: tenstriker):
[~jerryshao]
"spark.executor.cores" is to tell spark AM to request no of vcores from Yarn
per container. I think spark AM makes correct decision when dynamic allocation
is off but when its on it ignores spark.executor.cores value. I think DRF has
nothing to do with this in my opinion. Following are AM logs from two different
runs.
Run 1:
spark.dynamicAllocation.enabled = true
spark.executor.instances = 6
spark.executor.cores = 5
Dynamic allocation = true
17/01/09 19:05:49 INFO yarn.YarnRMClient: Registering the ApplicationMaster
17/01/09 19:05:49 INFO yarn.YarnAllocator: Will request 6 executor containers,
each with 1 cores and 11000 MB memory including 1000 MB overhead
17/01/09 19:05:49 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:1, disks:0.0>)
17/01/09 19:05:49 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:1, disks:0.0>)
17/01/09 19:05:49 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:1, disks:0.0>)
17/01/09 19:05:49 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:1, disks:0.0>)
17/01/09 19:05:49 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:1, disks:0.0>)
17/01/09 19:05:49 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:1, disks:0.0>)
Run 2:
spark.dynamicAllocation.enabled = false
spark.executor.instances = 6
spark.executor.cores = 5
17/01/09 19:01:39 INFO yarn.YarnAllocator: Will request 6 executor containers,
each with 5 cores and 11000 MB memory including 1000 MB overhead
17/01/09 19:01:39 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:5, disks:0.0>)
17/01/09 19:01:39 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:5, disks:0.0>)
17/01/09 19:01:39 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:5, disks:0.0>)
17/01/09 19:01:39 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:5, disks:0.0>)
17/01/09 19:01:39 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:5, disks:0.0>)
17/01/09 19:01:39 INFO yarn.YarnAllocator: Container request (host: Any,
capability: <memory:11000, vCores:5, disks:0.0>)
I can verify same fact via Spark UI when job is running that with dynamic
allocation there is only 1 task running per executor.
> Dynamic Resource Allocation not respecting spark.executor.cores
> ---------------------------------------------------------------
>
> Key: SPARK-19090
> URL: https://issues.apache.org/jira/browse/SPARK-19090
> Project: Spark
> Issue Type: Bug
> Affects Versions: 1.5.2, 1.6.1, 2.0.1
> Reporter: nirav patel
>
> When enabling dynamic scheduling with yarn I see that all executors are using
> only 1 core even if I specify "spark.executor.cores" to 6. If dynamic
> scheduling is disabled then each executors will have 6 cores. i.e. it
> respects "spark.executor.cores". I have tested this against spark 1.5 . I
> think it will be the same behavior with 2.x as well.
--
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
