CMK on Cluster Multi Node environment

[adpempx]

Hi,

I have a question very important: I have a Cluster Multi Node environment and I am using the following manifests in order to setup my CMK cluster:

• https://github.com/intel/CPU-Manager-for-Kubernetes/blob/master/resources/pods/cmk-init-pod.yaml
• https://github.com/intel/CPU-Manager-for-Kubernetes/blob/master/resources/pods/cmk-discover-pod.yaml
• https://github.com/intel/CPU-Manager-for-Kubernetes/blob/master/resources/pods/cmk-install-pod.yaml
• https://github.com/intel/CPU-Manager-for-Kubernetes/blob/master/resources/pods/cmk-reconcile-daemonset.yaml

Unfortunately I can't use cluster-init, because in that way the Helm chart will not have any information about the deployed POD and the problem here is for example, delete CMK Helm chart will not remove any resources created by python code.

Therefore I am using the second solution.
But I found CMK Init / Discover / Install PODs is running just on 1 single node and CMK Reconcile POD running on all nodes.
My questions are:

(1) if CMK Init / Discover / Install POD is running on just 1 single node, how can decide on which node I can deploy an application which use CMK?
(2) if CMK Init / Discover / Install POD runs on just 1 single node in a multi node environment, who made decision about the node? Which node CMK Init / Discover / Install POD, will CMK use?

For example, in my cluster multi node I have:

• cmk-discover Pod -> 172.20.55.65
• cmk-init Pod -> 172.20.55.67
• cmk-install Pod -> 172.20.55.65
• cmk-reconcile Pod -> 172.20.55.78, 172.20.55.65, 172.20.55.67

but in reality I have isolated CPU just on 172.20.55.67, so what I am expecting is to run all PODs just on that node.

Please let me know as those question are very important for us.

Thanks

[adpempx]

Hi,

I resolved by using DaemonSet + NodeAffinity for all CMK operations: discover, init, install and reconcile. It is great!

I am wondering one question: why CMK init sets CMK node's labels/ and node's capacity/allocatable resources related to the cores also in the nodes where there is not any CPU isolation?

I think that is a bug, as CMK should sets node's labels/ and node's capacity/allocatable resources only on nodes where CPU isolation is enabled.

Please let me know.

Thanks

[adpempx]

Hi,

do you have any news?

Thanks!

[pbrownlow7]

Hi @adpempx,

Apologies for the delay, I hadn't received any notifications.
I'm glad to see you've resolved the issues you were having, are there any others that I can help with?

As for CMK setting the labels and capacity/allocatable for nodes without CPU isolation, I'll have a look into that issue to determine if it is in fact a bug.

Thanks,
Philip

[adpempx]

Hi Philip,

I think it is a bug issue.

Please let me know,
Thanks!

[adpempx]

Hi,

do you have any news?

Thanks