The Stochastic-Learning-Based Deployment Scheme for Service Function Chain in Access Network

The research on topology-aware deployment for service function chain (SFC) is very important to a network slice technique. Nevertheless, most of the existing schemes on topology-aware deployment assume that the network topology information (NTI) is completely observed, which is unrealistic, considering the topology observation errors in practical network environment. In this paper, we consider the SFC deployment based on realistic topology sensing in fifth-generation cloud-radio access network (C-RAN). Due to the unavoidable errors, the realistic topology observation results merely represent partial NTI. Therefore, the partial observation Markov decision process (POMDP) is used in this paper to estimate the whole real topology condition. Then, a POMDP-based SFC deployment scheme is proposed. In this scheme, considering the particularity of SFC deployment in C-RAN, the SFC deployment problem is defined as a series of deployment decisions, including repair decisions, selection decisions, and resource allocation decisions. Our objective is to maximize the utility associated with the total delay and server-repair cost. And the POMDP scheme, according to the queue state information and partially observable NTI, makes deployment policies to maximize the utility by Bellman iteration. To reduce the iteration complexity, a point-based mingled heuristic value iteration algorithm is formulated in this paper. The simulation results show that the performance of C-RAN in terms of the system total delay and throughput can be significantly improved by using the proposed POMDP scheme.