Deep Q-Learning for Load Balancing Traffic in SDN Networks

Abstract

Load balancing is a widely used technique that aims to enable large network topologies, most commonly found in large data centers, to handle the constantly varying load of service requests. Traditional networks built on multi-vendor hardware and software present significant difficulties in the efficient and flexible application of load balancing techniques. Usually, solutions rely on high cost dedicated hardware and thus are used only for a subset of the tasks, resulting to limited flexibility for network administrators. Software Defined Networking (SDN) is a relatively new approach that enables flexible network management solutions to a number of problems, including that of efficient load-balancing. The key characteristics of decoupled centralized network control combined with programmability, allows the seamless integration of AI techniques to network management. Toward this direction, this paper employs deep reinforcement learning to effectively load balance requests to services in a data center network, resulting to an approach that is able to dynamically adapt to varying request loads, including changes in the infrastructure’s capabilities. The proposed approach is experimentally evaluated in order to support its feasibility, with very promising results.