| In recent years,network middleboxes,as the cornerstone of the Internet,support rapid development of a variety of internet services.However,middleboxes' shortcomings such as expensive price,high energy costs,less flexiblity and short lifecycle make it difficult for large-scale deployment and management.To solve aforementioned inconveniences,a new network architecture framework called Network Function Virtualization(NFV)is proposed,and become more and more popular under the support of service providers.NFV is now in the stage of rapid development,so there are still many challenges.The dynamic resource allocation and scheduling problem(Dynamic NFV-RA),as well as the bottleneck of the virtual network functions(VNFB),are the two most important challenges,which have been widely concerned by the academia and industry.Traditional NFV-RA algorithms are offline optimization algorithms,whose precondition is often assumed that there are service requests in advance,so Orchestrator has enough time to get the global optimal solution through the optimization algorithms and then completes the optimal placement.But in real situation,service requsets are not known in advance.Instead,they arrive to the system dynamically.So Orchestrator needs to handle incoming requests immediately,and completes placement in a limited period of time.However few algorithms,which are aiming to solving dynamic NFV-RA,cause some misssing feasible solutions because these algorithms narrow the search domain in order to improve the processing speed when selecting candidate nodes.Since the performance of VNFs can't be compared to the physical middleboxes,only a single or a small number of VNFs will cause the single point of failure and bottleneck.To solve the VNFB,the traditional horizontal expansion makes the overall system be very complex and difficult to detect errors.In this paper,we focus on the dynamic NFV-RA and VNFB problems and the main work is as follows:(1)Dynamic NFV-RA scheme based on resource sharing.The lifecycle of each service chain can be a short term,or a long term and the application time of chains of different tenants are often crisscross.So we first design a resource sharing model which allows a single VM running mutiple VNFs and these VNFs run in sequence.Based on this model,service requests can be embeded after waiting for a certain amount of time.It can effectively utilize the limited resources in the data center,and improve the comprehensive income.Second,in order to improve the resource allocation efficiency of the online algorithm,while not affecting the acceptance ratio of the request,we design a automatic monitoring mechanism based on the Finite State Machine(FSM),which can convert original active monitoring to a passive monitoring and reduce the overhead of resource monitoring.Finally,in order to reduce the switching overhead of scheduling as possible,we design and adopt a fast switching mechanism which defines a Linux namespace as a minimal allocation unit.So this is more lightweight and can reduce switching delay.The experimental results show that this mechanism can not only guarantee the running efficiency,also the acceptance ratio.(2)Lightweight network function virtualization system based on the pipeline technology.We find that the virtual switches deployed on each physical node can provide the functionality of the multi-level flow tables(Pipeline).Some simple network functions can be achieved though the combination of different flow tables.So based on the pipeline technology,we design a novel lightweight network function virtualization system,which provides simple and general APIs for most VNFs.Original VNFs can invoke general APIs by JSONRPC to realize the direct control of local pipelines,then inject partial functions into local pipelines,and finally avoid all network traffic being guided into reomte VNFs.The experimental results show this mechanism can reduce the load of remote VNFs and network links because partial network functions can be completed in local pipeline. |
PDF Full Text Request