Feature Overview (aka. Goal Summary)  

Modern CPU have multiple last level caches - clever alignment of workload with available cores can improve performance significantly. k8s is adding this capability to the CPUManager, and MicroShift should allow and support to utilize this.  

Goals (aka. expected user outcomes)

Support the prefer-align-cpus-by-uncorecache CPUManager feature on MicroShift.

Requirements (aka. Acceptance Criteria):

1. users must be able to activate the prefer-align-cpus-by-uncorecache policy via MicroShift configuration. 
2. users must receive same SLA for prefer-align-cpus-by-uncorecache  as it it is in OCP (TechPreview, GA). 

Questions to Answer (Optional):

1. How do we handle alpha/beta features behind a feature gate correctly? prefer-align-cpus-by-uncorecache is a k8s alpha/beta feature. We need a sustainable and holistic way to handle feature gates. This currently seems to be a bit of tech debt that needs to be fixed (see linked USHIFT-2813). 
   --> Moved that question to OCPSTRAT-2342
2. [Added Sep 30]How does this interact with Workload Partitioning? Any potential conflicts? Expecation / Hope is that it works seamlessly. 

Out of Scope

1. actual testing of the LLC alignment on AMD CPUS. We assume that this is done upstream/OpenShift wise. 

Background

k8s docs  

Upstream KEP: https://github.com/kubernetes/enhancements/blob/master/keps/sig-node/4800-cpumanager-split-uncorecache/README.md#summary 

Promotion PR to Beta in k8s 1.34: https://github.com/kubernetes/enhancements/pull/5390

OCP Tech Preview support statement: https://docs.google.com/presentation/d/1o_vniu1wT2C34I0q8RfGUx15YnRYGfi4cSk7k0dnOK8/edit?slide=id.g3209373ca5a_2_0#slide=id.g3209373ca5a_2_0 

Customer Considerations

Required for latency sensetive workload on AMD CPUs by industrial and telco customers. 

Documentation Considerations

This needs to be documented in the Configuring low latency section of the docs.

Interoperability Considerations

none