KubePACS: Kubernetes Cluster Using Performant, Highly Available, and Cost Efficient Spot Instances

Cloud users have long faced a trade-off when using spot instances: steep discounts come with reliability risks due to potential interruptions. Existing solutions like Karpenter, SpotVerse, and SpotKube primarily focus on cost or availability, but ignore performance variation across hardware types. KubePACS, developed by researchers at Korea University and Universitat Rovira i Virgili, tackles this by formulating node selection as a multi-objective optimization problem. It incorporates real-time data including spot prices, performance benchmarks, and a multi-node Spot Placement Score (SPS) to ensure high availability.

KubePACS solves this optimization efficiently using Integer Linear Programming (ILP) guided by the Golden Section Search (GSS) algorithm. It integrates with the Karpenter node autoscaler to jointly optimize instance-type selection and scaling decisions. A novel heuristic scales performance metrics for specialized instances, supporting workload-specific preferences. In evaluations across synthetic and real-world workloads, KubePACS achieved an average 55.09% and up to 81.06% higher performance per dollar compared to state-of-the-art solutions.