[ 
https://issues.apache.org/jira/browse/MESOS-5017?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15364414#comment-15364414
 ] 

Guangya Liu commented on MESOS-5017:
------------------------------------

I posted a patch here https://reviews.apache.org/r/49694/ , but found that the 
performance does not improve much with a benchmark test, [~bmahler] and 
[~jvanremoortere] , can you please help check and show your comments if any?

> Don't consider agents without allocatable resources in the allocator
> --------------------------------------------------------------------
>
>                 Key: MESOS-5017
>                 URL: https://issues.apache.org/jira/browse/MESOS-5017
>             Project: Mesos
>          Issue Type: Improvement
>          Components: allocation
>            Reporter: Dario Rexin
>            Assignee: Guangya Liu
>            Priority: Minor
>
> During the review r/43668/ , it come out an enhancement that if an agent has 
> not allocatable resources, the allocator should filter them out at the 
> beginning.
> {quote}
> Joris Van Remoortere Posted 1 month ago (March 16, 2016, 5:04 a.m.)
> Should we filter out slaves that have no allocatable resources?
> If we do, let's make sure we note that we want to pass the original slaveids 
> to the deallocate function
>  The issue has been resolved. Show all issues
> Dario Rexin 4 weeks ago (March 23, 2016, 4:25 a.m.)
> I'm not sure if it would be a big improvement. Calculating the available 
> resources if somewhat expensive and we have to do it again in the loop and 
> most slaves will probably have resources available anyway. The reason it's an 
> improvement in the loop is, that after we offer the resources to a framework, 
> we can be sure that they are all unavailable to the following frameworks 
> under the same role.
> Klaus Ma 4 weeks ago (March 23, 2016, 11:13 a.m.)
> @joris/dario, I think the improvement dependent on the workload patten: 1.) 
> for short running tasks, it maybe serveral tasks finished during the 
> allocation interval, so maybe no improvement; 2.) but for long running tasks, 
> slave/agent should be fully used in most of time, it'll be a big improvement. 
> I used to log MESOS-4986 to add a filter after stage 1 (Quota), but maybe 
> useless after revocable by default.
> Joris Van Remoortere 3 weeks, 6 days ago (March 23, 2016, 8:59 p.m.)
> Can you open a JIRA to consider doing this. Along Klaus' example, I'm not 
> convinced this wouldn't have a large impact in certain scenarios.
> {quote}



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to