CPU centric pod eviction similar to memory centric eviction

[suryadu]

No description provided.

[suryadu]

The main feature request here is to provide the mechanism in the safe scheduler that will dynamically monitor and balance the cluster through eviction of pods from busy nodes.

[mossuchida]

Currently, Kubernetes monitors resource utilization of each nodes. In case of out of memory, Kubernetes evicts offending pod from the node to release memory to bring back the proper performance of each nodes, however, when the CPU is fully utilized, CPU is compressed & throttled which is causing performance degradation of the application. We also need eviction mechanism for CPU similar to "out of memory" case. Since Kubernetes does not provide this option, we need this mechanism to avoid CPU performance degradation issue.