Research On Provisioning Of Service Function Chains In Heterogeneous NFV Environments

In order to cope with the rapid development of the Internet and to accommodate tremendous emerging applications,service providers have switched to network func-tion virtualization(NFV)instead of relying on special-purpose middleboxes,which can realize network appliactions with virtual network functions(vNFs)running on general-purpose platforms.Then,service providers can quickly realize various service compo-sitions with vNF service chains(vNF-SCs).However,previous studies mostly overlook the inherent differences among the substrate network elements and assume that all the vNFs are instantiated over the same type of platforms(e.g.,virtual machines).Since this assumption ignores the fact that vNFs have been realized on various software/hardware platforms and different network services may have different quality-of-service require-ments,it becomes impractical.Therefore,this paper studies the provisioning of service function chains in a heterogeneous NFV environment,which can instantiate vNFs on virtual machines,containers,and SmartNICs,to satisfy various quality-of-service de-mands while taking flexibility and cost-effectiveness into account.Firstly,we conduct experiments to explore performance of different NFV plat-forms and analyse the necessity of heterogeneous NFV environments.We build an experimental testbed that supports virtual machines,containers,and SmartNICs,and then measure the throughput and latency of traffic processing and resource usage of four types of vNFs implemented on them.After analyzing results,we find that each platform has its pros and cons.Then,we perform two experimental use cases to demonstrate that provisioning service function chains in a heterogeneous environment is feasible and can combine the advantages of software and hardware NFV platforms.In the meanwhile,it can enhance performance,flexibility and economy.Secondly,based on experimen-tal results,we formulate an optimization model and propose a heuristic algorithm to solve the problem.We leverage important parameters collected by experiments and formulate the problem as an integer linear programming(ILP)model which minimizes the total deployment cost to obtain the optimal solution.To greatly reduce time com-plexity,a heuristic algorithm is proposed.Simulation results show that the algorithm can effectively reduce cost and outperforms the existing heuristic.Finally,we design a polynomial-time approximation algorithm based on linear programming relaxation to provide near-optimal solutions whose gaps to the exact ones are bounded.The ap-proximation algorithm is proposed for the fact that heuristics do not have performance guarantee,and non-polynomial-time optimal models are not scalable.Then we the-oretically prove the approximation ratio and conduct extensive simulations including comparative algorithms to evaluate our proposal.The results confirm that with im-proved time-efficiency,our proposed algorithm provides near-optimal solutions whose approximation ratios are bounded.Although taking slightly longer time than the pro-posed heuristic,the approximation algorithm can obtain better deployment cost within polynomial time.In the meanwhile,it outperforms the existing approximation algo-rithm,which further verifies the effectiveness of our proposal.