hadoop-common-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Steve Loughran (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HADOOP-12855) Add option to disable JVMPauseMonitor across services
Date Tue, 08 Mar 2016 10:24:40 GMT

    [ https://issues.apache.org/jira/browse/HADOOP-12855?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15184747#comment-15184747

Steve Loughran commented on HADOOP-12855:

OK, so now the situation is
-the first JvmMonitor instance starts the monitor thread
-the last instance stops it.

This is going to be complicated enough to need a test. If some flag could let callers know
if the monitor thread was running, then a unit test coudl start two instances, stop the first
one, verify that the thread was running, stop the second, verify it was now stopped

also set monitorThread==null in the stop operation.

> Add option to disable JVMPauseMonitor across services
> -----------------------------------------------------
>                 Key: HADOOP-12855
>                 URL: https://issues.apache.org/jira/browse/HADOOP-12855
>             Project: Hadoop Common
>          Issue Type: Improvement
>          Components: performance, test
>    Affects Versions: 2.8.0
>         Environment: JVMs with miniHDFS and miniYarn clusters
>            Reporter: Steve Loughran
>            Assignee: John Zhuge
>         Attachments: HADOOP-12855-001.patch, HADOOP-12855-002.patch, HADOOP-12855-003.patch
> Now that the YARN and HDFS services automatically start a JVM pause monitor, if you start
up the mini HDFS and YARN clusters, with history server, you are spinning off 5 + threads,
all looking for JVM pauses, all printing things out when it happens.
> We do not need these monitors in minicluster testing; they merely add load and noise
to tests.
> Rather than retrofit new options everywhere, how about having a "jvm.pause.monitor.enabled"
flag (default true), which, when set, starts off the monitor thread.
> That way, the existing code is unchanged, there is always a JVM pause monitor for the
various services —it just isn't spinning up threads.

This message was sent by Atlassian JIRA

View raw message