/node-optimisation-demo

Apache License 2.0Apache-2.0

Node Optimisation Demo

The open source Telemetry-Aware Scheduling (TAS) tool can be used to optimise how Pods are scheduled onto Nodes in an energy-efficient manner.

TAS uses a metrics pipeline, Policies, and labeling strategies, which are currently the following:

  • scheduleonmetric
  • dontschedule
  • deschedule
  • labeling

TAS can be used to schedule Pods on Nodes based on various metrics such as a Node’s CPU utilization. According to "Energy-Proportional Computing: A new Definition", the optimal CPU utilization is 70-80% on average.

The following manifest can be used with TAS:

scheduleonmetric:
  rules:
  - metricname: CPUUtilization
    operator: GreaterThan
dontschedule:
  rules:
  - metricname: CPUUtilization
    operator: GreaterThan
    target: 0.8
deschedule:
  rules:
  - metricname: CPUUtilization
    operator: GreaterThan
    target: 0.8

This is saying to TAS:

  • To schedule a Pod, look at the CPUUtilization for each node, sort the values in descending order and pick a Node. The Node with capacity that is the closest to 70% will be chosen. If there is no capacity on the Node, the Horizontal Pod autoscaler should reschedule this Pod to a different Node.
  • New Pods won’t be scheduled on any Nodes where CPUUtilization is > 80%.
  • If you have existing pods that are on Nodes with CPUUtlization > 80%, mark them for descheduling. Then, we need the descheduler to actually deschedule the workloads.