Joint Autoscaling of Containers and Virtual Machines for Cost Optimization in Container Clusters

Abstract

Autoscaling enables container cluster orchestrators to automatically adjust computational resources, such as containers and Virtual Machines (VMs), to handle fluctuating workloads effectively. This adaptation can involve modifying the amount of resources (horizontal scaling) or adjusting their computational capacity (vertical scaling). The motivation for our work stems from the limitations of previous autoscaling approaches, which are either partial (scaling containers or VMs, but not both) or excessively complex to be used in real systems. This complexity arises from their use of models with a large number of variables and the addressing of two simultaneous challenges: achieving the optimal deployment for a single scheduling window and managing the transition between successive scheduling windows. We propose an Integer Linear Programming (ILP) model to address the challenge of autoscaling containers and VMs jointly, both horizontally and vertically, to minimize deployment costs. This model is designed to be used with predictive autoscalers and be solved in a reasonable time, even for large clusters. To this end, improvements and reasonable simplifications with respect to previous models have been carried out to drastically reduce the size of the resource allocation problem. Furthermore, the proposed model provides an enhanced representation of system performance in comparison to previous approaches. A tool called Conlloovia has been developed to implement this model. To evaluate its performance, we have conducted a comprehensive assessment, comparing it with two heuristic allocators with different problem sizes. Our findings indicate that Conlloovia consistently demonstrates lower deployment costs in a significant number of cases. Conlloovia has also been evaluated with a real application, using synthetic and real workload traces, as well as different scheduling windows, with deployment costs approximately 20% lower than heuristic allocators.