Thanks Sunil for the great work on this feature.
I looked through the design document, reviewed the code, and tested out branch
YARN-5881. The design makes sense and the code looks like it is implementing
the desing in a sensible way. However, I have encountered a couple of bugs. I
opened https://issues.apache.org/jira/browse/YARN-7575 to track my findings.
Basically, here's a summary:
The design document from YARN-5881 says that for max-capacity:
3) For each queue, we require: a) if max-resource not set, it automatically
set to parent.max-resource
When I try not setting
anyyarn.scheduler.capacity.<queue-path>.maximum-capacity, the RMUI scheduler
page refuses to render. It looks like it's in
CapacitySchedulerPage$LeafQueueInfoBlock.
Also... A job will run in the leaf queue with no max capacity set and it will
grow to the max capacity of the cluster, but if I add resources to the node,
the job won't grow any more even though it has pending resources.
Thanks,Eric
From: Sunil G <[email protected]>
To: "[email protected]" <[email protected]>; Hadoop Common
<[email protected]>; Hdfs-dev <[email protected]>;
"[email protected]" <[email protected]>
Sent: Friday, November 24, 2017 11:49 AM
Subject: [DISCUSS] Merge Absolute resource configuration support in Capacity
Scheduler (YARN-5881) to trunk
Hi All,
We would like to bring up the discussion of merging “absolute min/max
resources support in capacity scheduler” branch (YARN-5881) [2] into trunk
in a few weeks. The goal is to get it in for Hadoop 3.1.
*Major work happened in this branch*
- YARN-6471. Support to add min/max resource configuration for a queue
- YARN-7332. Compute effectiveCapacity per each resource vector
- YARN-7411. Inter-Queue preemption's computeFixpointAllocation need to
handle absolute resources.
*Regarding design details*
Please refer [1] for detailed design document.
*Regarding to testing:*
We did extensive tests for the feature in the last couple of months.
Comparing to latest trunk.
- For SLS benchmark: We didn't see observable performance gap from
simulated test based on 8K nodes SLS traces (1 PB memory). We got 3k+
containers allocated per second.
- For microbenchmark: We use performance test cases added by YARN 6775, it
did not show much performance regression comparing to trunk.
*YARN-5881* <https://issues.apache.org/jira/browse/YARN-5881>
#ResourceTypes = 2. Avg of fastest 20: 55294.52
#ResourceTypes = 2. Avg of fastest 20: 55401.66
*trunk*
#ResourceTypes = 2. Avg of fastest 20: 55865.92
#ResourceTypes = 2. Avg of fastest 20: 55096.418
*Regarding to API stability:*
All newly added @Public APIs are @Unstable.
Documentation jira [3] could help to provide detailed configuration
details. This feature works from end-to-end and we are running this in our
development cluster for last couple of months and undergone good amount of
testing. Branch code is run against trunk and tracked via [4].
We would love to get your thoughts before opening a voting thread.
Special thanks to a team of folks who worked hard and contributed towards
this efforts including design discussion / patch / reviews, etc.: Wangda
Tan, Vinod Kumar Vavilappali, Rohith Sharma K S.
[1] :
https://issues.apache.org/jira/secure/attachment/12855984/YARN-5881.Support.Absolute.Min.Max.Resource.In.Capacity.Scheduler.design-doc.v1.pdf
[2] : https://issues.apache.org/jira/browse/YARN-5881
[3] : https://issues.apache.org/jira/browse/YARN-7533
[4] : https://issues.apache.org/jira/browse/YARN-7510
Thanks,
Sunil G and Wangda Tan
