loadvariationriskbalancing

LoadVariationRiskBalancing Plugin

The LoadVariationRiskBalancing plugin is one of the Trimaran scheduler plugins and is described in detail in Trimaran: Real Load Aware Scheduling. The Trimaran plugins employ the load-watcher in order to collect measurements from the nodes as described here.

The (normalized) risk of a node is defined as a combined measure of the average and standard deviation of the node utilization. It is given by

risk = [ average + margin * stDev^{1/sensitivity} ] / 2

where average​ and stDev​ are the fractional (between 0 and 1) measured average utilization and standard deviation of the utilization over a period of time, respectively. The two parameters: margin​ and sensitivity​, impact the amount of risk due to load variation. In order to magnify the impact of low variations, the stDev​ quantity is raised to a fractional power with the sensitivity​ parameter being the root power. And, the margin​ parameter scales the variation quantity. The recommended values for the margin​ and sensitivity​ parameters are 1 and 2, respectively. Each of the two added terms is bounded between 0 and 1. Then, the divisor 2 is used to normalize risk between 0 and 1.

(Since the additional load due to the pod, that is the subject of scheduling, is not known in advance, we assume that its average and standard deviation load are the requested amount and zero, respectively.)

Risk is calculated independently for the CPU and memory resources on the node. Let worstRisk be the maximum of the two calculated risks. The score of the node, assuming that minScore is 0, is then computed as

score = maxScore * (1 - worstRisk)

Thus, the LoadVariationRiskBalancing plugin has the following configuration parameters:

• safeVarianceMargin : Multiplier (non-negative floating point) of standard deviation. (Default 1)
• safeVarianceSensitivity : Root power (non-negative floating point) of standard deviation. (Default 1)

In addition, we have the watcherAddress or metricProviderconfiguration parameters, depending on whether the load-watcher is in service or library mode, respectively.

Following is an example scheduler configuration with the LoadVariationRiskBalancing plugin enabled, and using the load-watcher in library mode, collecting measurements from the Prometheus server.

apiVersion: kubescheduler.config.k8s.io/v1beta3
kind: KubeSchedulerConfiguration
leaderElection:
  leaderElect: false
profiles:
- schedulerName: trimaran
  plugins:
    score:
      enabled:
       - name: LoadVariationRiskBalancing
  pluginConfig:
  - name: LoadVariationRiskBalancing
    args:
      safeVarianceMargin: 1
      safeVarianceSensitivity: 2
      metricProvider:
        type: Prometheus
        address: http://prometheus-k8s.monitoring.svc.cluster.local:9090