Research On Control And Optimization For Data Center Services Over Optical Networks

With the rapid development of cloud computing and data center, data center traffic is becoming the majority in the network, driving the evolution of underlying optical networks. Data center traffic requires large capacity,low latency and elasticity. In order to meet this, elastic optical network is seen as a significant candidate. Meanwhile, driven by novel data center services, control plane is required to be flexible, intelligent and open. How to realize the control and optimization for data center services is a key research topic.This dissertation investigates the optical control mechanism and optimization strategy for data center services from the following three aspects: interconnection control technique, resource allocation and optimization, and control plane safety. Research works are introduced as below:Firstly, with the purpose of interconnection for data center optical networks, based on software defined networks, this dissertation proposes a centralized virtual data center (VDC) provisioning architecture. The architecture introduces virtualization layer, which can provide virtual data center resources for tenants, e.g., computing, network topology, storage and optical bandwidth resources. Based on the architecture, the dissertation designs the interactive process, extends the OpenFlow based control protocol, proposes 2 IT resource allocation strategies, and sets up a physical node based testbed. Wireshark captures verify the feasibility of the architecture. In addition, for commerical optical transport network(OTN), this dissertation proposes a multi-domain multi-vendor centralized unified control plane architecture, extends CVNI based control protocol,designs interactive test cases between control plane and OTN devices,performs inter-continent field experiments, analyzes the communication protocol and collects OTN creation/deletion latency.Secondly, considering the problem of resource allocation and control optimization, this dissertation investigates advance reservation (AR)requests based on elastic optical networks, and establishes joint frequency and time domain resource model. In light of this, integer linear programming (ILP) is formulated to minimize spectrum utilization under static provisioning scenario. Scheduling methods are proposed and evaluated under 6-node network and NSFNet. Numerical results show that the proposed scheduling methods can decrease the capacity blocking probability. In dynamic provisioning scenario, in order to mitigate the spectral fragmentation caused by the dynamic requests, this dissertation proposes the time spectrum continuity (TSC) as the metric of the spectral fragmentation state. Based on TSC, MTLT, MHLT and MPT algorithms are proposed to mitigate the spectrum fragments. Simulation results indicate that TSC based algorithms achieve better performance of blocking probability compared with baselines. Considering some AR requests (e.g.,financial request) with high service level agreement (SLA), this dissertation investigates how to optimize the spectrum with protection. The dissertation proposes shared protection model in the joint frequency and time domains. Based on this model, ILP formulations are designed to minimize the resource usage. Meanwhile, for the scalability of the model,a greedy algorithm is proposed to maximize spectrum sharing. Simulation results show that the resource utilization efficiency of greedy algorithm is 19% higher than that of 1 + 1 protection.Thirdly, to ensure the stability of control plane, this dissertation proposes control plane intrusion detection based on tenant behavior analysis. For independent behavior, the dissertation designs a statistical control plane intrusion detection method, which establishes the probability model and probability threshold algorithm of tenant behaviors. Routing length and required bandwidth are used as features of the tenant behavior.This dissertation leverages a large number of training data to train the probability function, then uses calibration data to calculate the threshold of probability, finally uses testing data to measure the detection accuracy,which is around 83%. For continuous behaviors, the dissertation designs sequential control plane intrusion detection method, establishes shift cumulative sum of the tenant behavior. Sequential intrusion detection method takes arrival request number per time unit as the feature.Simulation results show that the detection accuracy is about 85%.