VNF Service Chain Deployment Problem Based On NUMA Architecture

With the development of computer network and the large-scale application of mobile Internet,there is a more diversified demand for network functions.Traditionally,the network function is implemented by hardware,which is low in flexibility,expansibility and ease of use,and cannot satisfy the needs of the rapid development of network function.The emergence of Network Function Virtualization(NFV)solves the problem well.NFV aims to use virtualization technology,in x86 and other general hardware software to achieve the original use of hardware functions,and through the decoupling of network functions,greatly improve the flexibility of network functions,while reducing expensive hardware updates and maintenance costs.With the increase of the number of cores in the server,many cores often need to access memory at the same time.But there is only one memory controller,and only one memory can be accessed by CPU at the same time.The traditional bus memory architecture has been unable to satisfy the multi-core computer's memory access demand.In order to solve the problem,the servers in the modern data center have adopted the Non Uniform Memory Access(NUMA)architecture.Compared with the deployment of virtual network function chain on traditional machines,the deployment of virtual network function chain on NUMA machines is more likely to cause resource bottleneck problems.This is mainly because not only the network card bandwidth will become the resource bottleneck,but also the memory capacity of nodes,the internal storage bandwidth of nodes,and the bandwidth of QPI bus between the different nodes on NUMA machines will appear bottlenecks.Therefore,when the virtual network function chain is deployed in the actual NUMA cluster,the resource utilization rate of the cluster will be extremely low.To solve the problem,this thesis studies how to improve the resource utilization of NUMA machines by considering the constraints of multiple resources,memory capacity of each NUMA node,memory bandwidth of each NUMA node,bandwidth of QPI bus between different NUMA nodes,number of CPU,bandwidth of network card,when deploying virtual network function chains on NUMA clusters.For the typical process of actual deployment,two typical scenarios are considered: online service deployment scenario and offline redeployment scenario.In the online service deployment scenario,the requests arrive in sequence,and the deployment can only be based on the requests at the current time,while the future requests may destroy the optimality of the current deployment,resulting in suboptimal results.For the optimization problem,we consider the offline redeployment scenario,which periodically redeploys all existing requests to achieve better deployment effect.(1)For the online service scenario,we have verified through practical experiments that the memory accessing the remote NUMA Node will significantly reduce throughput and resource utilization.To solve the problem,we analyze the resource competition on NUMA server and the data flow of virtual network function chain in NUMA.We model the problem of improving resource utilization by considering memory capacity,memory bandwidth,QPI bandwidth,total CPU and network card bandwidth to integer linear programming.Since the requests arrive one by one,a new chain will arrive after the deployment of the current chain,so when we deploy the currently arrived virtual network function chain,we cannot run out of network bandwidth resources.In order to accommodate as many virtual network functional chains as possible,it is necessary to avoid the fragmentation of cluster resources,including hardware resources and network resources.If there are only a few resources left in each server,but the resources needed by the virtual network function chain that is about to arrive are larger than the remaining resources of each server,then the virtual network function chain cannot be deployed.If the previous deployment reduces the fragmentation as much as possible,the virtual network function chain may be able to be deployed in the cluster.According to the actual test results,we design a heuristic virtual network function chain deployment algorithm.The algorithm deploys the adjacent virtual network functions in the service chain to the same NUMA node as much as possible,so as to avoid the throughput drop caused by memory access to remote NUMA Node.A large number of simulations show that our online algorithm can improve resource utilization 9%-17% compared to first fit algorithm?(2)For VNF chain batch deployment scenario,we use a Monte Carlo Tree Search algorithm to solve the problem.For the sub optimal deployment scheme in online deployment,we further improve resource utilization through batch deployment.The main challenge is to determine both the deployment of virtual network functions and the redistribution of traffic.We use the Monte Carlo Tree Search algorithm to solve the problem of probability and statistics.In each iteration,we first determine the deployment of virtual network function,and then get the optimal traffic distribution scheme by solving the integer linear programming.In order to avoid falling into the local optimum in each iteration,both the optimal results obtained by the historical iteration and certain randomness are considered in the decision-making process.Finally,according to the deployment location of each virtual network function in the virtual network function service chain,a decision tree is constructed,each node represents a placement,and the child node represents the placement of the next virtual network function.By traversing all the leaf nodes,the optimal deployment scheme is determined by comparing the values of leaf nodes.A large number of simulations show that the performance of the designed algorithm is improved to 17%-39%?This thesis investigates the deployment of VNF service chains in NUMA architecture server clusters and considers two typical scenarios for the timing of requests: online service scenarios and batch deployment scenarios.In different scenarios,the problem is modeled and formally described separately,and the VNF service chain deployment algorithm based on greedy policy and the Monte Carlo tree search-based service chain deployment algorithm are proposed,which,after a lot of experiments,shows a big improvement over the traditional algorithm.