Q-Learning-Based Workload Consolidation for Data Centers With Composable Architecture

Abstract

Composable or disaggregated architectures have emerged as a solution to address the drawbacks of server-based architectures in data centers, such as resource inefficiency and limited scalability. This article focuses on the workload consolidation problem where we aim to consolidate workloads that spread over many underutilized (resource) nodes onto fewer ones, with the two objectives of minimizing the number of active nodes and workload migrations, thereby enhancing energy efficiency and resource utilization. To address this problem, we propose a Q-learning-based reinforcement learning method that yields an approximate Pareto front, providing a set of (approximate) optimal solutions catering to different preferences for the two objectives. The performance of the proposed method is validated by comparing it to integer linear programming, simulated annealing, first fit, and first fit decreasing algorithms.