[
https://issues.apache.org/jira/browse/YUNIKORN-2069?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17779561#comment-17779561
]
Craig Condit edited comment on YUNIKORN-2069 at 10/25/23 4:03 PM:
------------------------------------------------------------------
Hi [~rrajesh] assigned you this one. I think the general approach should be:
# Create two queues, one with a very low guaranteed limit (500m CPU), one with
a very high guaranteed limit (much larger than a node size, say 5000 CPU)
# Select a schedulable node from the cluster
# Schedule a number of small, low priority pause tasks of 500m CPU in the
low-guarantee queue (enough to fill the selected node), with a node selector
that only allows placement on that node. Wait for all tasks to become running.
Since a node may not have exactly a multiple of 500m available, round down when
calculating number of tasks needed.
# Schedule a larger (1000m CPU) high-priority task in the high-guarantee
queue, also with the same node selector.
# Wait (perhaps up to 60 seconds) for the larger task to schedule. This should
require several of the low-priority tasks to be preempted.
# Shutdown and cleanup all tasks created.
was (Author: ccondit):
Hi [~rrajesh] assigned you this one. I think the general approach should be:
# Create two queues, one with a very low guaranteed limit (500m CPU), one with
a very high guaranteed limit (much larger than a node size, say 5000 CPU)
# Select a schedulable node from the cluster
# Schedule a number of small, low priority pause tasks of 500m CPU in the
low-guarantee queue (enough to fill the selected node), with a node selector
that only allows placement on that node. Wait for all tasks to become running.
# Schedule a larger (1000m CPU) high-priority task in the high-guarantee
queue, also with the same node selector.
# Wait (perhaps up to 60 seconds) for the larger task to schedule. This should
require several of the low-priority tasks to be preempted.
# Shutdown and cleanup all tasks created.
> Add e2e test covering YUNIKORN-2068
> -----------------------------------
>
> Key: YUNIKORN-2069
> URL: https://issues.apache.org/jira/browse/YUNIKORN-2069
> Project: Apache YuniKorn
> Issue Type: Test
> Reporter: Craig Condit
> Assignee: Rajesh Kanhaiya Lal
> Priority: Major
>
> YUNIKORN-2068 added a deadlock fix during preemption - we need to create a
> test case that can preempt tasks from a specific node (probably using node
> selectors).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
