μOpt: An Efficient Optimal Autoscaler for Microservice Applications
Authors:
- 25 September 2023
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 67-76
Abstract
Microservices are a popular architecture for cloudbased applications subject to stringent performance requirements. To effectively serve variable workloads, autoscaling allocates computational resources ideally at the lowest possible cost. Although several autoscaling techniques have already been proposed in the literature, they suffer from high computational complexity. Here we propose μOpt as a computationally efficient model-based autoscaler for microservices. By solving a nonlinear optimization problem that embeds a layered queueing network (LQN) model, μOpt computes optimal configurations maximizing performance while minimizing allocated resources. We validate μOpt on a benchmark microservice application, reporting fast solution times (~10 -1 s) that enable prompt reactions to highly variable workloads. Compared to a state-of-the-art autoscaler based on LQN and genetic algorithms, μOpt achieves higher performance (~6%-8%) with significantly fewer allocated resources (~15%-35%) in the presence of both synthetic and real-world workloads.Keywords
This publication has 59 references indexed in Scilit:
- Applying reinforcement learning towards automating resource allocation and application scalability in the cloudConcurrency and Computation: Practice and Experience, 2012
- Performance Prediction of Service-Oriented Systems with Layered Queueing Networks2010
- The weighted sum method for multi-objective optimization: new insightsStructural and Multidisciplinary Optimization, 2009
- Large-scale nonlinear programming using IPOPT: An integrating framework for enterprise-wide dynamic optimizationComputers & Chemical Engineering, 2009
- Interaction tree algorithms to extract effective architecture and layered performance models from tracesJournal of Systems and Software, 2007
- The interior-point revolution in optimization: History, recent developments, and lasting consequencesBulletin of the American Mathematical Society, 2004
- Applying the UML Performance Profile: Graph Grammar-Based Derivation of LQN Models from UML Specifications2002
- A generalized processor sharing approach to flow control in integrated services networks: the single-node caseIEEE/ACM Transactions on Networking, 1993
- Model predictive control: Theory and practice—A surveyAutomatica, 1989
- Solutions of ordinary differential equations as limits of pure jump markov processesJournal of Applied Probability, 1970