Many-core Based Programmable Virtual Router Data Plane Design And Implementation

The next generation network becomes to a hot topic because of t he existingInternet is lack of scalability, security, and high-performance. However, the nextgeneration network is not compatible with the existing Internet. Therefore, theinnovation of future network needs a testbed which surpports multiple differentnetworks at the same time. Programmale virtual router(PEARL) not only solves theproblem of the coexistence of multiple networks, but also meets requirements of thefuture network, such as dynamically scalable, flexible and programmable,high-performance and other characteristics. As the core equipment of networks, theperformance of routers is critical. Modern high-performance routers generally adoptthe structure which the control plane and the data plane are separated. And many–corenetwork processors are widely used in the data plane. Many-core network processorsnot only have lots of cores executing programs parallelly, but also have some assistantprocessors which have been specially optimized for network packet processing.Existing task partition method is not suitable for router system, and existing coreallocation method has a short of core resource utilization, therefor e, the overallperformance can be improved by some scheme. Against these problems, the mainwork of this paper includes2points.First, this paper improves the flexibility of packet processing by task partition.Reasonable task partition can improve the flexibility and parallel of packet processing.Packet processing is partitioned into3major modules including Parse, Lookup,Forward, and2auxiliary modules including Deliver and Control. Cores are assignedto these modules and they can run in parrallel. This method not only improved theperformance, but also improved flexibility. Users can add modules dynamically, suchas firewall, IDS, and so on, through remote protocol.Second, this paper proposes a core resource optimizing method. Against theshortcoming of EQUI, proposes a resource optimizing allocation method which isbased on mathematical expectation of network traffic in many core systems.Implementation includes3steps. Step1, packets are divided into two planes, onebelongs to the control plane, and the other belongs to the data plane. In a time period t,the probability of each kind of packets can be statisticed by some networkmeasurement methods. Step2, each type of packets has different procedure, and thus situation results modules have different workload. Total CPU clocks of each moduleneeds can be calculated according to probability obtained in step1. Step3, cores areassigned proportionally to each module according the probability of each module’stotal CPU workload obtained in the step2. And experiment results show that thisoptimal scheme improves20%in minimum-size packets forwarding rate comparedwith EQUI.