Adaptive Parallel Processing Mechanism For Service Function Chains In NFV

The network function virtualization(NFV)that decouples network functions(NFs)from proprietary equipment and implements them as virtual network functions(VNFs)on general-purpose servers,provides an efficient solution to reduce the deployment costs and enable flexible NF placement and service customization.The NFV has the advantages of high flexibility,low cost,and efficient maintenance and deployment to networks.In the NFV,network services are provided as a series of predefined VNFs called service function chain(SFC).Traditionally,the VNFs in an SFC are processed one by one in a specific order,which represents the serial processing mode.The SFC delays in serial processing mode will increase linearly with the lengths of SFCs,which can cause unacceptable delay in the low-latency applications.This paper mainly studies the parallel SFC(PSFC)processing mechanism include PSFC optimized deployment and PSFC scheduling.Firstly,the order dependency between NFs is studied and parallelized VNFs are separated from the original chain and placed into the branch of parallel processing to generate the PSFC.Then a PSFC deployment method(Para Match)based on resource matching is proposed to realize the optimized deployment of PSFC.Simulation experiments show that the algorithm can effective reduce the SFC delay by 45%.Secondly,we propose an adaptive PSFC processing mechanism(APPM)based on reinforcement learning method to further optimize the processing efficiency of PSFC and solve the PSFC scheduling problem.Specifically,the APPM proposes a parallelism optimization algorithm(POA)to adjust the structures of PSFCs.The POA can eliminate unnecessary parallelism of the PSFC and effectively reduce extra resource consumption.Further,a joint optimization algorithm based on reinforcement learning(Jo RL)is employed to jointly deploy and schedule PSFCs,which avoids the VNF scheduling after SFC deploying and breaking the delay balance in sub-SFCs of a PSFC.The numerical results show that APPM can significantly improve the performance of the SFC parallel processing,reduce the total SFC delay,and increase both of the network resource utilization and service acceptance rate.This thesis will provide a key technical basis for ultra-low latency services deploying and scheduling,and have a far-reaching practical significance for building a fast and efficient future network.