Optimizing the Cost of Executing Mixed Interactive and Batch Workloads on Transient VMs

Abstract

Container Orchestration Platforms (COPs), such as Kubernetes, are increasingly used to manage large-scale clusters by automating resource allocation between applications encapsulated in containers. Increasingly, the resources underlying COPs are virtual machines (VMs) dynamically acquired from cloud platforms. COPs may choose from many different types of VMs offered by cloud platforms, which differ in their cost, performance, and availability. While transient VMs cost significantly less than on-demand VMs, platforms may revoke them at any time, causing them to become unavailable. While transient VMs' price is attractive, their unreliability is a problem for COPs designed to support mixed workloads composed of, not only delay-tolerant batch jobs, but also long-lived interactive services with high availability requirements. To address the problem, we design TR-Kubernetes, a COP that optimizes the cost of executing mixed interactive and batch workloads on cloud platforms using transient VMs. To do so, TR-Kubernetes enforces arbitrary availability requirements specified by interactive services despite transient VM unavailability by acquiring many more transient VMs than necessary most of the time, which it then leverages to opportunistically execute batch jobs when excess resources are available. When cloud platforms revoke transient VMs, TR-Kubernetes relies on existing Kubernetes functions to internally revoke resources from batch jobs to maintain interactive services' availability requirements. We show that TR-Kubernetes requires minimal extensions to Kubernetes, and is capable of lowering the cost (by 53%) and improving the availability (99.999%) of a representative interactive/batch workload on Amazon EC2 when using transient compared to on-demand VMs.