Basic workflow for PEAKS scheduler-plugin

[knarayan]

Below is a basic workflow for PEAKS scheduler-plugin:

• Model the relationship between CPU utilization and the energy consumption for each node of the cluster.
• For a Pod to be scheduled, find the change in the utilization of a node if the pod is to be scheduled on that node.
  • Correspondingly, find the change in the energy consumption on that node.
  • This step requires estimating the Pod resource needs.
    • Considering Pod resource limits alternatively to start.
• Repeat the above step to collect the change in the energy consumption for each node.
• Choose the node with least change in the energy consumption.

CC: @atantawi and @husky-parul

[knarayan]

Details on power model training and inferencing:

• On each cluster node, we run the benchmark workloads and collect the metrics <CPU utilization, power consumption>.
• We train a ML based power model for each node using the above metrics data.
• Inside the PEAKS plugin, we import the power model parameters for each node.
  • If the power models are re-trained subsequently and the parameters change, updates to the parameters are passed to the plugin (say via a webhook)
• For each pod to be scheduled, the plugin estimates its resource usage.
  • For each node, the plugin computes change in the node utilization using the estimated pod resource usage and the node utilization
  • Plugin also computes respective “change in the power consumption” for each node by carrying out power model inferencing
• Node with lower “change in the power consumption” is preferred for Pod placement by the PEAKS plugin

CC: @husky-parul and @atantawi