Self-adaptive, Requirements-driven Autoscaling of Microservices

TL;DR

This paper introduces MS-RA, a self-adaptive, requirements-driven autoscaling solution for microservices that outperforms standard methods by efficiently allocating resources based on service-level objectives.

Contribution

The paper presents MS-RA, a novel self-adaptive autoscaling approach utilizing SLOs and the MAPE-K loop, improving resource efficiency and SLO satisfaction in microservices.

Findings

MS-RA reduces CPU and memory usage significantly.

MS-RA requires fewer replicas than HPA.

MS-RA maintains SLO compliance effectively.

Abstract

Microservices architecture offers various benefits, including granularity, flexibility, and scalability. A crucial feature of this architecture is the ability to autoscale microservices, i.e., adjust the number of replicas and/or manage resources. Several autoscaling solutions already exist. Nonetheless, when employed for diverse microservices compositions, current solutions may exhibit suboptimal resource allocations, either exceeding the actual requirements or falling short. This can in turn lead to unbalanced environments, downtime, and undesirable infrastructure costs. We propose MS-RA, a self-adaptive, requirements-driven solution for microservices autoscaling. MS-RA utilizes service-level objectives (SLOs) for real-time decision making. Our solution, which is customizable to specific needs and costs, facilitates a more efficient allocation of resources by precisely using the right amount to meet the defined requirements. We have developed MS-RA based on the MAPE-K self-adaptive loop, and have evaluated it using an open-source microservice-based application. Our results indicate that MS-RA considerably outperforms the horizontal pod autoscaler (HPA), the industry-standard Kubernetes autoscaling mechanism. It achieves this by using fewer resources while still ensuring the satisfaction of the SLOs of interest. Specifically, MS-RA meets the SLO requirements of our case-study system, requiring at least 50% less CPU time, 87% less memory, and 90% fewer replicas compared to the HPA.