hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Wangda Tan (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (YARN-6599) Support rich placement constraints in scheduler
Date Wed, 10 Jan 2018 03:59:00 GMT

    [ https://issues.apache.org/jira/browse/YARN-6599?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16319662#comment-16319662

Wangda Tan commented on YARN-6599:

Discussed with [~asuresh] and [~kkaranasos] today, a summary of what we discussed.

1) To support check cardinality for intra-app/inter-app. We will reuse the TargetExpression.targetKey
to specify application label. Application label is a new concept which could be: 

- {{APPLICATION_LABEL/SELF}}: to check cardinality within the app.
- {{APPLICATION_LABEL/<APPLICATION_ID>}}: to check cardinality under app with given
app id. 
- {{APPLICATION_LABEL/<LABEL>}}: to check cardinality under app with given label. 

What we will immediately add is {{APPLICATION_LABEL/SELF}}, and move all other staffs to future
when we want to do cross-application affinity/anti-affinity. 

2) SELF target-type will be removed in YARN-7709. Application needs to copy source tags to
target tags when wants to affinity/anti-affinity to itself.

3) Will keep node partition related logic in the patch for reviews. In the future after YARN-3409,
we need to add all node-related properties to node attributes, such as {{yarn.io/host.name/<host-name>}},
{{yarn.io.failure.domain/<domain-name>}}, {{yarn.io.node.partition/<partition-name>}}.
Reason of doing this is, we can support all node related attribute in the same logic, no need
to do if checks like: 

if (targetKey == node_partition) {
    // check node.getPartition

4) Since we have two approaches: 
- Placement processor. (Added by YARN-7612) 
- AppPlacementAllocator for SchedulingRequest (added by this patch). 
Both approaches have merits, the first one can consider multiple-apps/multiple-scheduling-requests
at the same time. And the latter one keeps same scheduler behavior   and support features
like node partition, allocation order under different priority/allocation-id/app-priority/queue-order

Since the whole scheduling request is a new feature, we will include two configs:
a. Enable placement processor (YARN-7612 already includes configs)
b. Enable scheduling request handled by app placement allocator.

Both a/b are disabled by default and cannot be enabled at the same time.

An alternative solution is to have one config which supports enum values like processor/scheduler.
(we need a better name for this). Thoughts here?

5) For Scheduler allocate interface, we will stick to approach introduced by this patch. Extracting
a common class for SchedulerRequest/ResourceRequest cannot simplify the problem and might
introduce compatibility issues. 

[~asuresh]/[~kkaranasos], please add if any missed/inaccurate.

> Support rich placement constraints in scheduler
> -----------------------------------------------
>                 Key: YARN-6599
>                 URL: https://issues.apache.org/jira/browse/YARN-6599
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>            Reporter: Wangda Tan
>            Assignee: Wangda Tan
>         Attachments: YARN-6599-YARN-6592.003.patch, YARN-6599-YARN-6592.004.patch, YARN-6599-YARN-6592.005.patch,
YARN-6599-YARN-6592.006.patch, YARN-6599-YARN-6592.007.patch, YARN-6599-YARN-6592.008.patch,
YARN-6599-YARN-6592.wip.002.patch, YARN-6599.poc.001.patch

This message was sent by Atlassian JIRA

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

View raw message