hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Carlo Curino (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (YARN-4195) Support of node-labels in the ReservationSystem "Plan"
Date Thu, 17 Dec 2015 06:04:46 GMT

    [ https://issues.apache.org/jira/browse/YARN-4195?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15061560#comment-15061560

Carlo Curino commented on YARN-4195:

The posted patch is a rebase of our working branch. It is useful to kickstart reviewing/discussion,
but probably not ready to commit as-is.

In general, the patch adds the support of tracking per-label resource allocations in the ReservationSystem
(and performs a bit of cleanups).

The one key point worth discussing is related to partitions vs labels:
 # to ensure that capacity per node label means anything both queues in the CS (and reservations
here) are forced to refer to "partitions" of the underlying nodes. 
 # this is annoying and limiting. For example if I have GPU and PUBLICIP as my desired user-visible
labels, I would have to manually express all 4 combinations GPU_PUBLICIP, GPU_NOT-PUBLICIP,
NOT-GPU_PUBLICIP, NOT-GPU_NOT-PUBLICIP when defining queues, and worst, the users will be
forced to express their needs in the same form. For example a user that only cares about running
on {{GPU}} would have to say: {{GPU_PUBLICIP OR GPU_NOT-PUBLICIP}}. 
 # we propose an improvement: 
 ## Internally the system tracks partitions. 
 ## administrators configuring queues do so at the partition level
 ## users are allowed to express their job needs in terms of labels (and the system internally
converts this into partitions)
 ## users can reserve in terms of labels (and the system internally converts this into partitions)

It is generally provable that an arbitrary expression of labels can be represented as an OR
of partitions (or in disjunctive-normal-form). In this and future patches we have a version
of this using a mix of JEXL and ad-hoc accelerations (JEXL was too slow in some of our tests).
However, I believe  [~chris.douglas] has a better version of this, which uses a cool algorithm
which skips the conversion to DNF. We should plug these here.

More generally, this improvement can be of general use for CS (and in the future FS) to expose
a nicer API to users.

(Once again, marking patch as ready to get the conversation going, some cleanup still required)


> Support of node-labels in the ReservationSystem "Plan"
> ------------------------------------------------------
>                 Key: YARN-4195
>                 URL: https://issues.apache.org/jira/browse/YARN-4195
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>            Reporter: Carlo Curino
>            Assignee: Carlo Curino
>         Attachments: YARN-4195.patch
> As part of YARN-4193 we need to enhance the InMemoryPlan (and related classes) to track
the per-label available resources, as well as the per-label
> reservation-allocations.

This message was sent by Atlassian JIRA

View raw message