Title:Heuristic and Reinforcement Learning Algorithms for Dynamic Service Placement on Mobile Edge Cloud

Abstract:Edge computing hosts applications close to the end users and enables low-latency real-time applications. Modern applications inturn have adopted the microservices architecture which composes applications as loosely coupled smaller components, or services. This complements edge computing infrastructure that are often resource constrained and may not handle monolithic applications. Instead, edge servers can independently deploy application service components, although at the cost of communication overheads. Consistently meeting application service level objectives while also optimizing application deployment (placement and migration of services) cost and communication overheads in mobile edge cloud environment is non-trivial. In this paper we propose and evaluate three dynamic placement strategies, two heuristic (greedy approximation based on set cover, and integer programming based optimization) and one learning-based algorithm. Their goal is to satisfy the application constraints, minimize infrastructure deployment cost, while ensuring availability of services to all clients and User Equipment (UE) in the network coverage area. The algorithms can be extended to any network topology and microservice based edge computing applications. For the experiments, we use the drone swarm navigation as a representative application for edge computing use cases. Since access to real-world physical testbed for such application is difficult, we demonstrate the efficacy of our algorithms as a simulation. We also contrast these algorithms with respect to placement quality, utilization of clusters, and level of determinism. Our evaluation not only shows that the learning-based algorithm provides solutions of better quality; it also provides interesting conclusions regarding when the (more traditional) heuristic algorithms are actually better suited.