Mailing-List: contact issues-help@aurora.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@aurora.apache.org
Date: Fri, 17 Feb 2017 16:18:41 +0000 (UTC)
From: "Mehrdad Nurolahzade (JIRA)" <jira@apache.org>
To: issues@aurora.apache.org
Message-ID: <JIRA.13043766.1487281236000.111225.1487348321711@Atlassian.JIRA>
In-Reply-To: <JIRA.13043766.1487281236000@Atlassian.JIRA>
References: <JIRA.13043766.1487281236000@Atlassian.JIRA> <JIRA.13043766.1487281236775@jira-lw-us.apache.org>
Subject: [jira] [Commented] (AURORA-1894) Inline preemption filter in
 PreemptionVictimFilterImpl
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit
archived-at: Fri, 17 Feb 2017 16:18:46 -0000


    [ https://issues.apache.org/jira/browse/AURORA-1894?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15872045#comment-15872045 ] 

Mehrdad Nurolahzade commented on AURORA-1894:
---------------------------------------------

I used YourKit (default settings), can't say if this caused by profiler bytecode manipulations for sampling or not. 

However, just by looking at this logic, I am not sure if compiler can do such optimizations. The Lambda is dependent on the method parameter ({{ITaskConfig pendingTask}}).

{code}
    private Predicate<PreemptionVictim> preemptionFilter(final ITaskConfig pendingTask) {
      return possibleVictim -> {
        boolean pendingIsPreemptible = tierManager.getTier(pendingTask).isPreemptible();
        boolean victimIsPreemptible =
            tierManager.getTier(possibleVictim.getConfig()).isPreemptible();

        if (!pendingIsPreemptible && victimIsPreemptible) {
          return true;
        } else if (pendingIsPreemptible == victimIsPreemptible) {
          // If preemptible flags are equal, preemption is based on priority within the same role.
          if (pendingTask.getJob().getRole().equals(possibleVictim.getRole())) {
            return pendingTask.getPriority() > possibleVictim.getPriority();
          } else {
            return false;
          }
        } else {
          return false;
        }
      };
    }
{code}

> Inline preemption filter in PreemptionVictimFilterImpl
> ------------------------------------------------------
>
>                 Key: AURORA-1894
>                 URL: https://issues.apache.org/jira/browse/AURORA-1894
>             Project: Aurora
>          Issue Type: Task
>          Components: Scheduler
>            Reporter: Mehrdad Nurolahzade
>            Priority: Minor
>              Labels: newbie
>
> Profiling preemption logic, I can see {{PreemptionVictimFilterImpl#preemptionFilter()}} is producing ~200K/sec lambda objects to be used by {{filterPreemptionVictims()}}:
> {code:title=PreemptionVictimFilterImpl.filterPreemptionVictims()}
>       FluentIterable<PreemptionVictim> preemptableTasks = FluentIterable.from(possibleVictims)
>           .filter(preemptionFilter(pendingTask));
> {code}
> Inline this logic (refactor to loop) to remove the need to create these short-lived lambda objects.


--
This message was sent by Atlassian JIRA
(v6.3.15#6346)