Resource Allocation In Software-defined Networking

Over recent decades,the scale of topologies and the number of applications in the network grow rapidly.Yet,existing IP network faces challenges on management due to the distributed control logic.Software-defined networking(SDN)is a novel paradigm that aims to solve these challenges by promoting the central control plane.The centralization of network control introduces the global views of the network states and resources.This thesis starts with addressing the challenges that how to guarantee the performance of SDN control actions on the switch.Furthermore,this thesis discusses the resource allocation challenges while adopting SDN in the data center and public clouds.(1)TCAM flow table allocation in SDN switchesThe programmatic control of SDN demands fast and frequent modification of the flow table implemented with Ternary Content-Addressable Memory(TCAM).However,the latency of inserting rules into the TCAM flow table is usually unpredictable.The experiments show that the unpredictable performance of Ternary Content-Addressable Memory(TCAM)can hurt the performance of applications by 40% – 60%.In view of this,this chapter presents the design and evaluation of a novel resource allocation framework that provides performance guarantees by partitioning and optimizing switch TCAM.The evaluations show that with less than 5% storage overheads,the frameworks can provide 5ms insertion latency guarantees that correspond to up to 80% improvement of the application level performance.(2)Computing resource allocation in SDN switchesWhile modifying the flow table,switch usually needs to parse the control messages prior to writing the rules into the flow table.This latency of the parsing operation is determined by the computing resources of the switch.Yet,the infrastructure-level resource isolation and performance guarantee on SDN switches are not supported by the existing network virtualization approaches.In view of this,this chapter proposes the Switch Resource Allocation Framework(SRAF).SRAF promises to provide tenants with the resource isolation and performance guarantees over the switches.The experiments show that SRAF can support multiple tenants with the performance guarantee,and only incur less than 4% overhead in throughput and 3% extra delay.(3)Topology resource allocation in SDN data centersBeyond switches,Data Center Networks(DCN)is a popular use case of SDN.Recently,the solutions to improve the operating efficiency and performance of DCN are widely investigated.However,while the current solutions achieve high optimization efficiency the network operators still need to manually configure the switch when the workloads and capacities change,this makes the configuration of network devices in the data center a difficult task.Meanwhile,according to existing works,this chapter argues that actually many of the existing DCN optimization methods share the same design and similar architecture.In view of this,this chapter presented a deep learning based method that aims at providing the intermediate representation of the network topology generally and further solving this class of DCN problems with self-driving technology.To demonstrate the benefit of the deep learning based approach,this chapter implemented,configured,and evaluated the self-driving framework that automatically augments the data center topology.The evaluations show that the proposed framework achieves nearoptimal results.(4)Allocating service function chain across multiple clouds and SDN data centerDespite the local data center and switches,the emerging of Network Function Virtualization allows the network operators to outsource the network functions to the public clouds in order to reduce the operational cost.Unfortunately,challenges of Quality of Service guarantee still exist while minimizing the operational cost with public clouds.This chapter addresses these challenges by discussing the deployment situation with consideration of a large number of public clouds,different pricing schemes,additional latencies,and the mapping between the Virtual Network Function(VNF)specification and its price.This chapter designed MOSC,a heuristic algorithm to allocate the Service Function Chain across multiple clouds based on Hidden Markov Model(HMM)and Viterbi algorithm.The evaluations show that MOSC saves up to 80% cost compared with traditional local network deployment method.Additionally,MOSC can achieve up to 50% cost savings compared with the result of the state-of-art optimization algorithm.