ImpalaE: Towards an optimal policy for efficient resource management at the edge

Abstract

Edge computing is an extension of cloud computing where physical servers are deployed closer to the users in order to reduce latency. Edge data centers face the challenge of serving a continuously increasing number of applications with a reduced capacity compared to traditional data center. This paper introduces ImpalaE, an agent based on Deep Reinforcement Learning that aims at optimizing the resource usage in edge data centers. First, it proposes modeling the problem as a Markov Decision Process, with two optimization objectives: reducing the number of physical servers used and maximize number of applications placed in the data center. Second, it introduces an agent based on Proximal Policy Optimization, for finding the optimal consolidation policy, and an asynchronous architecture with multiple workers-shared learner that enables for faster convergence, even with reduced amount of data. We show the potential in a simulated edge data center scenario with different VM sizes based on Microsoft Azure real traces, considering CPU, memory, disk and network requirements. Experiments show that ImpalaE effectively increases the number of VMs that can be placed per episode and that it quickly converges to an optimal policy.