Tuning a Kubernetes Horizontal Pod Autoscaler for Meeting Performance and Load Demands in Cloud Deployments

Abstract

In the context of scaling a business-critical medical service that involves electronic medical record storage deployed in Kubernetes clusters, this research addresses the need to optimize the configuration parameters of horizontal pod autoscalers for maintaining the required performance and system load constraints. The maximum entropy principle was used for calculating a load profile to satisfy workload constraints. By observing the fluctuations in the existing workload and applying a kernel estimator to smooth its trends, we propose a methodology for calculating the threshold parameter of a maximum number of pods managed by individual autoscalers. The results obtained indicate significant computing resource savings compared to autoscalers operating without predefined constraints. The proposed optimization method enables significant savings in computational resource utilization during peak loads in systems managed by Kubernetes. For the investigated case study, applying the calculated vector of maximum pod count parameter values for individual autoscalers resulted in about a 15% reduction in the number of instantiated nodes. The findings of this study provide valuable insights for efficiently scaling services while meeting performance demands, thus minimizing resource consumption when deploying to computing clouds. The results enhance our comprehension of resource optimization strategies within cloud-based microservice architectures, transcending the confines of specific domains or geographical locations.

SLA-Adaptive Threshold Adjustment for a Kubernetes Horizontal Pod Autoscaler

“…One of the latest works that aims to dynamically optimise the performance of the HPA and address resource waste avoidance is [21]. Augustyn et al [21] suggest an approach to identify the maximum number of pods to be provisioned by the HPA, which allows customers to continue to use the HPA while improving resource utilisation. The work does not aim to ensure system performance conforms with the [5] SLO.…”