Distributedness based scheduling

TL;DR

This paper introduces a novel Kubernetes scheduling strategy that enhances resource distribution by considering usage patterns and distributedness, leading to improved resource utilization and balance across cluster nodes.

Contribution

It proposes a label-driven, distributedness-based scheduling enhancement for Kubernetes that optimizes resource balance considering usage patterns and supports redistribution.

Findings

Improved resource balance across nodes.

Enhanced scheduler considers usage patterns for placement.

Supports redistribution to optimize existing pod placement.

Abstract

Efficient utilization of computing resources in a Kubernetes cluster is often constrained by the uneven distribution of pods with similar usage patterns. This paper presents a novel scheduling strategy designed to optimize the distributedness of Kubernetes resources based on their usage magnitude and patterns across CPU, memory, network, and storage. By categorizing resource usage into labels such as "cpu high spike" or "memory medium always," and applying these to deployed pods, the system calculates the variance or distributedness factor of similar resource types across cluster nodes. A lower variance indicates a more balanced distribution. The Kubernetes scheduler is enhanced to consider this factor during scheduling decisions, placing new pods on nodes that minimize resource clustering. Furthermore, the approach supports redistribution of existing pods through simulated scheduling to improve balance. This method is adaptable at the cluster, namespace, or application level and is integrated within the standard Kubernetes scheduler, providing a scalable, label-driven mechanism to improve overall resource efficiency in cloud-native environments.