Development of network services embedding method using reinforcement learning

Abstract

In the last few years, the Network Functions Virtualization (NFV), a network architecture approach, has become essential for all the services provider companies. With NFV architectures, providers can reduce the requirements for specialized hardware [1], which may stay unused for most of the time if it serves only a few requests. But in order to use most of the cloud infrastructure, they require methods for mapping a service onto the virtualized infrastructure. There’s where Network Service Embedding comes into play, to help providers optimize the distribution of the physical resources to fulfill the customers’ needs as fast as possible and in a more reliable way. Network Service Embedding [2] (NSE) methods can take into account more complex needs that a client may specify, such as low latency, and bandwidth limits except for CPU or memory demands. Also, NSE helps providers to manage their resources efficiently, therefore, serving as many clients in a given period of time, is giving them the ability to increase their profits. This is also important for the clients as they can experience the quality of service and lower costs based on their needs. The purpose of this Master’s thesis is to develop a method for the optimized embedding of network services onto a virtualized infrastructure (e.g., data center) using supportive learning techniques based on Reinforcement Learning algorithms, as opposed to heuristic methods that are mostly employed. For the implementation of this work, Python [3] was used as programming language, the DRL models developed using Tensorflow [4] framework and the generated service graph were created with NetworkX [5] framework.