Minimizing active nodes in MEC environments: A distributed learning-driven framework for application placement

Abstract

Application placement in Multi-Access Edge Computing (MEC) must adhere to service level agreements (SLAs), minimize energy consumption, and optimize metrics based on specific service requirements. In distributed MEC system environments, the placement problem also requires consideration of various types of applications with different entry distribution rates and requirements, and the incorporation of varying numbers of hosts to enable the development of a scalable system. One possible way to achieve these objectives is to minimize the number of active nodes in order to avoid resource fragmentation and unnecessary energy consumption. This paper presents a Distributed Deep Reinforcement Learning-based Capacity-Aware Application Placement (DDRL-CAAP) approach aimed at reducing the number of active nodes in a multi-MEC system scenario that is managed by several orchestrators. Internet of Things (IoT) and Extended Reality (XR) applications are considered in order to evaluate close-to-real-world environments via simulation and on a real testbed. The proposed design is scalable for different numbers of nodes, MEC systems, and vertical applications. The performance results show that DDRL-CAAP achieves an average improvement of 98.3% in inference time compared with the benchmark Integer Linear Programming (ILP) algorithm, and a mean reduction of 4.35% in power consumption compared with a Random Selection (RS) algorithm.