Performance Modeling and Capacity Planning for SAAS Applications

Abstract

Providing quality services in cloud-based systems is a critical factor. SaaS based applications typically experience dynamic workloads which compels system architects and providers to look for ways to keep their application services scalable, but finding the right number of services can be a cumbersome task if they solely rely on testing and maintaining the application in lab environments. This is primarily due to the significant time and costs involved in setting up the system. In this thesis, we propose a lightweight method to perform capacity planning of the applications. This approach combines an analytical tool that models the system as a closed network of queues and utilizes a numerically stable algorithm, SMVA, to approximate performance metrics, and a simulation framework developed to capture more intricacies of our underlying system environment and platform. We validate the proposed methods on a microservices app deployed on a Kubernetes cluster that captures key metrics like throughput, response time, and pod CPU utilization. The results show acceptable agreement among the SMVA predictions, simulation outputs, and empirical observations, therefore confirming the effectiveness of our approach.