What is it?

A load-aware scheduler plugin which scores the nodes based on the risk of overloading resources. An acceptable overload is specified by the administrator via two configuration parameters: SafeUtilization and SafePercentile. For example, a setting of 0.95 and 0.10, respectively, for the CPU resource, means that a CPU utilization of more than 95% during up to 10% of the time is acceptable. In this case, this LoadRiskAwareCommitment scheduler plugin will evaluate the chance (risk) of a node having CPU utilization above 95%. This risk is compared to the acceptable percentile and a score is calculated. A low (high) score implies high (low) risk, with a zero score corresponding to a risk higher than the acceptable risk.

Why we need it?

Typically, the values of requested (and limit) amount of resources for a pod do not reflect the actual resource utilization during the lifetime of the pod. A desirable resource management criterion might be to place pods, not only based on requested resources, but also to avoid overloading resources. A simplistic threshold value for overload is not enough, as it does not consider the chance of being above that threshold. Thus, one needs to specify both a SafeUtilization and a SafePercentile. As a result, pods will be less likely to be assigned to risky nodes. This will avoid nodes being saturated, leading to bad performance. Also, it lowers the probability of evictions. Further, this LoadRiskAwareCommitment scheduler plugin will accommodate scenarios of over allocating resources.

How its done?

• The cluster administrator configures two parameters: SafeUtilization and SafePercentile.  (Default values might be 0.90 and 0.10, respectively.) We consider CPU and memory as the resources of interest. The parameters might be resource-specific.
• A load monitor (Prometheus or a load watcher) calculates the average utilization and standard deviation (std) quantities for the resources of interest over a period of time
• The scheduler plugin fits a Beta probability distribution to the observed average and standard deviation values using the method of matching moments (simple two algebraic equations). This is done for all resources of interest.
• Risk is calculated as the probability the utilization is above the SafeUtilization value (tail of the distribution).
• Relative risk is the value of risk divided by the SafePercentile, and upper bounded by one.
• Each resource will have a score value, calculated as: ( 1 - relativeRisk) * maxScore.
• The node score is the minimum (worst case) score among the resources of interest.

An example is provided.