Machine Learning Based Approach for Multi- domain Service Deployment in Future Netorks

Abstract

The capabilities of Network Function Virtualization (NFV) and Software Defined Networking (SDN) will be exploited by Network Service Providers (NSPs) in order for them to fulfil the di vergent and stringent requirements, that are characteristic of the Fifth Generation (5G) network and in essence realise a dynamic and flexible communication network. With the softwarization paradigm, Network Functions (NFs) will be decoupled from the underlying customized mid dle ware and virtualized into Virtual Network Functions (VNFs). These VNFs, when tailored and connected together in an ordered manner will constitute a Service Function Chain (SFC) which will satisfy a consumer’s requirements. However, due to reliability and availability re lated issues, this SFC may be served by more than one Infrastructure Provider (InP) and thus subjected to different traffic flow and policies more so in an environment of limited topology information disclosure. In the event that the SFC is to be deployed across multiple domains, it will present challenges in terms of how to effectively map the VNFs to their respective resource constrained NFs in the underlying infrastructure whilst maintaining their order of instantia tion. In this project, an efficient Reinforcement Learning (RL) algorithm that exploits a request decomposition technique and learns from experiential intelligence is proposed. This algorithm coordinates the partitioning of the SFCs into sub chains and the multi- layer graph mapping of VNFs onto the shared infrastructure so as to embed requests of any topology. The key results obtained showed that the proposed algorithm realised an improvement of up to 26% in terms of acceptance ratio and an improvement of up to 10% in terms of provisioning cost.