[
https://issues.apache.org/jira/browse/YUNIKORN-772?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17392895#comment-17392895
]
Chen Yu Teng edited comment on YUNIKORN-772 at 8/4/21, 9:06 AM:
----------------------------------------------------------------
Hi [~wilfreds] and [~wwei]
As [~wilfreds] saying, default scheduler scheduling a pod is same as the doc.
If YK merges YUNIKORN-706, YK scheduling behavior will be same as default
scheduler.
Default schedule and YK record node resource usage via same calculation.
But there are some information discrepancy in YK-web.
Thanks [~wwei] for summary .
Effective requirement from pod may be not actual resource requirement in whole
pod lifecycle
So i consider that we add a variable to allocationAsk to store actual usage
in current pod lifecycle.
Let user know there are some allocated resource may be not used in running
state.
I don't think YK needs to modify user's yaml to improve usage problem
automatically.
If users want to improve usage in allocated resource. related information can
help them.
# Fit doc implement in default scheduler and its logic is also used in
internal cache
[https://github.com/kubernetes/kubernetes/blob/master/pkg/scheduler/framework/plugins/noderesources/fit.go#L135-L179
[https://github.com/kubernetes/kubernetes/blob/master/pkg/scheduler/framework/types.go#L583]|https://github.com/kubernetes/kubernetes/blob/master/pkg/scheduler/framework/plugins/noderesources/fit.go#L135-L179]
# kubelet find next container to start and maintains pod via kuberuntime
[https://github.com/kubernetes/kubernetes/blob/master/pkg/kubelet/kuberuntime/kuberuntime_manager.go#L538]
was (Author: yukali):
Hi [~wilfreds] and [~wwei]
As [~wilfreds] saying, default scheduler scheduling a pod is same as the doc.
If YK merges YUNIKORN-706, YK scheduling behavior will be same as default
scheduler.
Default schedule and YK record node resource usage via same calculation.
But there are some information discrepancy in YK-web.
Thanks [~wwei] for summary .
Effective requirement from pod may be not actual resource requirement in pod
lifecycle
So i consider that we add a variable to allocationAsk to store actual usage
in current pod lifecycle.
Let user know there are some allocated resource may be not used in running
state.
I don't think YK needs to modify user's yaml to improve usage problem
automatically.
If users want to improve usage in allocated resource. related information can
help them.
# Fit doc implement in default scheduler and its logic is also used in
internal cache
[https://github.com/kubernetes/kubernetes/blob/master/pkg/scheduler/framework/plugins/noderesources/fit.go#L135-L179
https://github.com/kubernetes/kubernetes/blob/master/pkg/scheduler/framework/types.go#L583|https://github.com/kubernetes/kubernetes/blob/master/pkg/scheduler/framework/plugins/noderesources/fit.go#L135-L179]
# kubelet find next container to start and maintains pod via kuberuntime
[https://github.com/kubernetes/kubernetes/blob/master/pkg/kubelet/kuberuntime/kuberuntime_manager.go#L538]
> Add AllocationAskUpdate in UpdateRequest
> ----------------------------------------
>
> Key: YUNIKORN-772
> URL: https://issues.apache.org/jira/browse/YUNIKORN-772
> Project: Apache YuniKorn
> Issue Type: Improvement
> Components: scheduler-interface
> Reporter: Chen Yu Teng
> Priority: Major
> Fix For: 1.0.0
>
> Attachments: description.png
>
>
> When resource requirement changes in pod phase, shim will update locally and
> core-scheduler doesn't know it.
> Web and core-scheduler shows resource requirement changed when pod phase
> transited from initialization to running state
> Any case?
> This situation happens in YUNIKORN-706 .
> Core-scheduler creates allocationAsk which is stable and calculated from
> containers.
> If initcontainer resource requirements replace some containers requirements,
> core-scheduler will create allocationAsk and see it as fixed allocation and
> ignore resource requirement changed when pod phase transited from
> initialization to running state.
> !description.png|width=693,height=328!
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
