Research On Resource Allocation And System Optimization For NFV

In recent years,Network Function Virtualization(NFV)has emerged as a promising network architecture paradigm for quick,flexible deployment of network services.By transforming the way of implementing network functions,NFV advocates replacing hardware appliances with software network functions(NFs)running on commercial-off-the-shelf severs,thereby simplifying the deployment and management of network services,and reducing capital expenditure and operational expenditure.With the development of NFV,NF service chains have been applied in telecom operator,data center and enterprise networksIn practical NFV deployment,NF instances are commonly running in virtual machines(VMs)or containers.As different NFs have different resource requirements,operators need to provision sufficient VMs and launch corresponding NFs for service chains when users demand for services,in order to provide the desired performance.For diverse traffic conditions and different goals,operators require to deploy NFs according to different strategies,while not harming the service chain performance.Therefore,for operators,how to properly allocate their resources and provide performance guarantee for users is of great importanceTo cope with the problems in NFV deployment,existing research on resource allocation is either reactive or proactive.Current reactive research assumes that traffic requests are known,and mainly achieves resource scheduling through a centralized scheme.However for telecom operators,the deployment of NFs across geo-distributed COs has a dual collaborative allocation from both NF to CO and NF to server,for load balancing the resources in each CO.Current proactive research considers the traffic fluctuations,but ignores the routing problem in service chains,or exploits an unreasonable approach to predict future resource requirements for each NF,or performs vertical scaling on the fly.Besides,for implementing high-performance NFV platforms,existing NFV frameworks either focus on standalone NFs or cannot automatically perform dynamic NF scaling.To deal with the limitations of existing research,we have researched resource allocation and system optimization problem in NFV with both theoretical analysis and system implementation,from the following three aspectsReactive-based NF assignment for load balancing of resources.Telecom operators need to consider the dual NF assignment problem mentioned above while deploying NF service chains in their geo-distributed COs.When user traffic requests are known,we propose a two-stage solution.First,we select an appropriate number of COs from all COs under the operators' overall budget,while minimizing the communication cost.This problem corresponds to finding the largest subgraph,which is proved to be NP-hard.We solve it with a 2-approximation algorithm.Second,we turn to a shadow-routing based approach for assigning NFs in service chains across the selected COs.It belongs to a virtual queueing system and brings theoretical basis for problems with various objectives and constraints.To solve the dual assignment problem,each routing decision is easily made by the values and updates of virtual queues,thus achieving load balancing of both computing and bandwidth resources across all selected COs.The algorithm runs continuously to adapt to changes in network demands,and is proven to be asymptotically optimalProactive-based NF scaling and flow routing with demand prediction.Considering the traffic fluctuations in practice,operators need to dynamically adjust NF deployments for their users in the cloud,and face the problem of either under-or over-provisioned resources To minimize the cost of operators,we propose a joint online algorithm to guide operators to dynamically deploy NFs.The algorithm first employs an efficient online learning method called follow the regularized leader(FTRL)to predict the upcoming flow rates.An upper bound of the regret of FTRL is then proven.Based on the predicted rates,the algorithm provisions new instances with the best or maximum capacity in advance for bottleneck NFs Next,we model the assignment of new instances as a variable-sized bin packing problem,and propose an efficient online heuristic with a competitive ratio of 3/2.For service chain routing,we propose an online primal-dual algorithm with a competitive ratio of(1+?).Finally,an overall ratio that approaches 3/2 over time is then derived to validate the good performance of the proposed algorithm,which jointly considers the regret of predicting flow rates and the competitiveness of new instance deploymentLoad-aware NF service chaining system with dynamic scaling.For NFV platforms,efficiently processing packets through service chains requires instantaneous regulatory mechanism based on traffic changes.We propose FlexNFV,an service chaining framework designed to provide automatic and efficient NF scaling capabilities for NFV platforms By capturing the required information from the underlying system at the runtime,such as per-packet processing cost of different NF s,traffic characteristics and queues between NF s,FlexNFV can accurately compute the current load of each NF.Based on the chain-level load,FlexNFV rapidly performs NF scaling,thereby avoiding performance degradation of service chains.Besides,FlexNFV also introduces optimizations such as power-aware management and elimination of queue conflicts.Compared to existing works,FlexNFV can flexibly adapt to real-time traffic demand,while not only improving throughput but also reducing packet drop rate.