Gradient Descent based Vertical Auto-Scaling Algorithm for Microservices

As cloud technology advances, microservice architectures are increasingly popular for handling large-scale applications. These applications are decomposed into multiple microservices that communicate with each other via APIs or RPCs, to improve their flexibility and reliability. However, different microservices require diverse resources, it would be a significant challenge to keep efficient resource scaling while ensuring the quality of services. Traditional scaling approaches include horizontal and vertical scaling ones, but the vertical scaling approach often risks service disruption. Kubernetes introduces an In-place Pod Vertical Scaling approach, which enables seamless scaling and presents fresh opportunities for optimizing resource usage. This study proposes a Gradient Descent based Vertical AutoScaling (GDVAS) algorithm, which utilizes the gradient descent approach to automatically optimize resource allocation based on the Quality of Service (QoS). In the experiment, we employ three gradient descent techniques: Adam, Momentum, and AdaGrad to improve the resource allocation. Experimental results show that the GDVAS could efficiently determine the optimal resource allocation within six iterations. Furthermore, GDVAS is particularly efficient when applying Momentum as the gradient descent technique, highlighting its suitability for optimizing resource allocation in microservice environments.