[
https://issues.apache.org/jira/browse/YARN-9209?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16754249#comment-16754249
]
Wangda Tan commented on YARN-9209:
----------------------------------
[~tarunparimi], [~cheersyang],
Actually, this is by design. There're two reasons which we don't want to
support ANY node partition.
1) The purpose of node partition is to separate cluster to multiple "virtual"
clusters which have independent queue capacities, ACLs, usages, etc. We want
user explicitly ask for one partition instead of "give me arbitrary partition".
And the typical reason to have multiple partitions is to isolate important
resources, such as GPU resources. We definitely don't want a sleep MR test job
use resources on GPU partition unless it is asked explicitly.
2) A more technical challenge is, once multiple partitions are asked, how to
calculate pending resource gonna be a problem. We don't want to double counting
pending resources. (because there's only one ask) But it also seems not correct
to only increase default partition. A miscalculated pending resource will cause
trouble when doing preemption.
We originally wanted to support ANY partition when we did original node
partition support (YARN-796), we gave up because of the two reasons. Please let
me know if there're any solutions to get #1 and #2 resolved.
> When nodePartition is not set in Placement Constraints, containers are
> allocated only in default partition
> ----------------------------------------------------------------------------------------------------------
>
> Key: YARN-9209
> URL: https://issues.apache.org/jira/browse/YARN-9209
> Project: Hadoop YARN
> Issue Type: Bug
> Components: capacity scheduler, scheduler
> Affects Versions: 3.1.0
> Reporter: Tarun Parimi
> Assignee: Tarun Parimi
> Priority: Major
> Attachments: YARN-9209.001.patch
>
>
> When application sets a placement constraint without specifying a
> nodePartition, the default partition is always chosen as the constraint when
> allocating containers. This can be a problem. when an application is
> submitted to a queue which has doesn't have enough capacity available on the
> default partition.
> This is a common scenario when node labels are configured for a particular
> queue. The below sample sleeper service cannot get even a single container
> allocated when it is submitted to a "labeled_queue", even though enough
> capacity is available on the label/partition configured for the queue. Only
> the AM container runs.
> {code:java}{
> "name": "sleeper-service",
> "version": "1.0.0",
> "queue": "labeled_queue",
> "components": [
> {
> "name": "sleeper",
> "number_of_containers": 2,
> "launch_command": "sleep 90000",
> "resource": {
> "cpus": 1,
> "memory": "4096"
> },
> "placement_policy": {
> "constraints": [
> {
> "type": "ANTI_AFFINITY",
> "scope": "NODE",
> "target_tags": [
> "sleeper"
> ]
> }
> ]
> }
> }
> ]
> }
> {code}
> It runs fine if I specify the node_partition explicitly in the constraints
> like below.
> {code:java}
> {
> "name": "sleeper-service",
> "version": "1.0.0",
> "queue": "labeled_queue",
> "components": [
> {
> "name": "sleeper",
> "number_of_containers": 2,
> "launch_command": "sleep 90000",
> "resource": {
> "cpus": 1,
> "memory": "4096"
> },
> "placement_policy": {
> "constraints": [
> {
> "type": "ANTI_AFFINITY",
> "scope": "NODE",
> "target_tags": [
> "sleeper"
> ],
> "node_partitions": [
> "label"
> ]
> }
> ]
> }
> }
> ]
> }
> {code}
> The problem seems to be because only the default partition "" is considered
> when node_partition constraint is not specified as seen in below RM log.
> {code:java}
> 2019-01-17 16:51:59,921 INFO placement.SingleConstraintAppPlacementAllocator
> (SingleConstraintAppPlacementAllocator.java:validateAndSetSchedulingRequest(367))
> - Successfully added SchedulingRequest to
> app=appattempt_1547734161165_0010_000001 targetAllocationTags=[sleeper].
> nodePartition=
> {code}
> However, I think it makes more sense to consider "*" or the
> {{default-node-label-expression}} of the queue if configured, when no
> node_partition is specified in the placement constraint. Since not specifying
> any node_partition should ideally mean we don't enforce placement constraints
> on any node_partition. However we are enforcing the default partition instead
> now.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
