[ 
https://issues.apache.org/jira/browse/YUNIKORN-807?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17407164#comment-17407164
 ] 

Chia-Ping Tsai commented on YUNIKORN-807:
-----------------------------------------

[~ccondit] Thanks for this nice patch and sorry for late comment.

It seems to me a small behavior change is introduced by this patch. This patch 
scores node according to `resources.LargestUsageShare`. However, the largest 
usage in our cluster is always storage (and all nodes have same storage 
capacity). It results in the order is always based on `node id`. Before this 
patch, the order will compare remaining "usage" (vcore, memory, etc) when 
largest usage is equal.

There are two possible solutions.
 # the nodeScore of nodeRef should keep all usages ([]float64) and the 
implementation of `Less` function should reuse `resources.CompareShares`
 # add an new config to enable users to control scope. for example, the 
following config means we should compare vcore instead of largest resource 

{code:java}
nodesortpolicy:
  type: binpacking
  usage: vcore
{code}

That’s just my two cents :)

> Improve performance of node sorting
> -----------------------------------
>
>                 Key: YUNIKORN-807
>                 URL: https://issues.apache.org/jira/browse/YUNIKORN-807
>             Project: Apache YuniKorn
>          Issue Type: New Feature
>          Components: core - scheduler
>            Reporter: Craig Condit
>            Assignee: Craig Condit
>            Priority: Major
>              Labels: pull-request-available
>             Fix For: 1.0.0
>
>         Attachments: Node Sorting Performance Improvement.pdf
>
>
> YuniKorn currently sorts all nodes on demand whenever scheduling of a 
> container occurs. This causes significant performance degradation as the 
> number of nodes increases.
> If we replace the on-demand sorting with a B-Tree sorted proactively, we can 
> improve performance considerably.
> This is a similar approach to YUNIKORN-21, but without the associated 
> behavioral changes.
> I've attached a design document with the details of the approach and the 
> performance improvement gained.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to