hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Karthik Kambatla (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (YARN-5035) FairScheduler: Adjust maxAssign dynamically when assignMultiple is turned on
Date Thu, 26 May 2016 05:50:13 GMT

     [ https://issues.apache.org/jira/browse/YARN-5035?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Karthik Kambatla updated YARN-5035:
-----------------------------------
    Attachment: yarn-5035-3.patch

Updated patch to use fitsIn. Note that fitsIn checks <= and not <. This leads to allocating
one container more than floor(available/2). This is okay - the goal here is to have a maxassign
that is (1) proportional to available resources, and (2) considers resources and not containers.


On making this ratio of 0.5 configurable, I considered it but decided against for two reasons:
# I don't anticipate users will change this. In that case, a boolean config is simpler than
a number that they have to think about. 
# {{1/2 + 1/4 + ...}} converges nicely and is somewhat easier to reason about than say {{1/3
+ 1/9 + ...}}. Also, if we really need a config for this, we could add one in the future.


> FairScheduler: Adjust maxAssign dynamically when assignMultiple is turned on
> ----------------------------------------------------------------------------
>
>                 Key: YARN-5035
>                 URL: https://issues.apache.org/jira/browse/YARN-5035
>             Project: Hadoop YARN
>          Issue Type: Improvement
>          Components: fairscheduler
>    Affects Versions: 2.8.0
>            Reporter: Karthik Kambatla
>            Assignee: Karthik Kambatla
>         Attachments: yarn-5035-1.patch, yarn-5035-2.patch, yarn-5035-3.patch
>
>
> When assignMultiple is turned on, today the number of containers allocated to node within
a single heartbeat is determined by maxAssign. Picking a higher value for this prevents the
spreading of the workload, and a smaller number leads to slow ramp up. 
> The proposal is to make it dynamic - half of the unallocated resources on the node. 



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

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org


Mime
View raw message