ignite-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Dmitriy Sorokin (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (IGNITE-6171) Native facility to control excessive GC pauses
Date Thu, 07 Dec 2017 10:29:00 GMT

    [ https://issues.apache.org/jira/browse/IGNITE-6171?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16281641#comment-16281641

Dmitriy Sorokin commented on IGNITE-6171:

[~avinogradov], review new patch, please!

> Native facility to control excessive GC pauses
> ----------------------------------------------
>                 Key: IGNITE-6171
>                 URL: https://issues.apache.org/jira/browse/IGNITE-6171
>             Project: Ignite
>          Issue Type: Task
>          Components: general
>    Affects Versions: 2.3
>            Reporter: Vladimir Ozerov
>            Assignee: Dmitriy Sorokin
>              Labels: iep-7, usability
>             Fix For: 2.4
> Ignite is Java-based application. If node experiences long GC pauses it may negatively
affect other nodes. We need to find a way to detect long GC pauses within the process and
trigger some actions in response, e.g. node stop. 
> This is a kind of Inception \[1\], when you need to understand that you sleep while sleeping.
As all Java threads are blocked on safepoint, we cannot use Java's thread to detect Java's
GC. Native threads should be used instead.
> Proposed solution:
> 1) Thread 1 should periodically call dummy JNI method returning current time, and set
this time to shared variable;
> 2) Thread 2 should periodically check that variable. If it has not been changed for some
time - most likely we are in GC pause. Once certain threashold is reached - trigger compensating
action, whether this is a warning, process kill, or what so ever.
> Justification: crossing native -> Java boundaries involves safepoints. This way Thread
1 will be trapped if STW pause is in progress. Java method cannot be empty, as JVM is smart
enough and can deduce it to no-op. 
> \[1\] http://www.imdb.com/title/tt1375666/

This message was sent by Atlassian JIRA

View raw message