hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jason Lowe (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (YARN-5393) [Umbrella] Optimize YARN tests runtime
Date Fri, 12 Aug 2016 21:00:22 GMT

    [ https://issues.apache.org/jira/browse/YARN-5393?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15419470#comment-15419470

Jason Lowe commented on YARN-5393:

A lot of this bloat is from over-sleeping.  It's sad to watch the CPU utilization during unit
tests because it's often atrociously low.  Doing some silly manual sampling with jstack while
some of the unit tests were running found a lot of tests waiting in one of the following methods:
- MockRM.waitForState
- MockAM.waitForState
- RMStateStoreTestBase.waitNotify

I'm sure I missed a lot of other notorious places, but the common theme is sleeping for far
too long.  One second is a really, really long time for a modern CPU, and if tests hammer
on methods that do this (and many tests hammer waitForState methods) then the seconds quickly
start to pile up into minutes.  I'd love to see most of the sleeps above 10 milliseconds reduced,
since I think that's where a lot of our runtime is going.

I noticed that YARN-2921 put in a minimum sleep time because reportedly some unit tests were
failing without it.  That minimum for a few tests is killing the performance on a lot of tests.
 We need to track down those racy tests and fix them rather than forcing a long sleep time
for any test that calls those waitForState methods.  The longer sleep isn't a real fix for
those racy tests anyway, rather it just reduces the frequency at which they fail.

> [Umbrella] Optimize YARN tests runtime 
> ---------------------------------------
>                 Key: YARN-5393
>                 URL: https://issues.apache.org/jira/browse/YARN-5393
>             Project: Hadoop YARN
>          Issue Type: Test
>            Reporter: Vinod Kumar Vavilapalli
> When I originally merged MAPREDUCE-279 into Hadoop, *all* of YARN tests used to take
10 mins with pretty good coverage.
> Now only TestRMRestart takes that much time - we'ven't been that great writing pointed
- short tests.
> Time for an initiative to optimize YARN tests. And even after that, if it takes too long,
we go the MAPREDUCE-670 route.

This message was sent by Atlassian JIRA

To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org

View raw message