Multi-stage Switching Fabric And Scheduling Algorithms For Research

Confronted with the explosion of communication traffic in Internet, the communication networks architects make great efforts to provide scalability for switch architecture in the current routers and switches. Single-stage switches are hard to implement in scalable switches for the limitation of pins and complexities of IC design. To build a larger switching fabric by using interconnection networks including several switch elements is commonly used as scalability solutions. The traditional scheduling algorithms which used in multi-stage networks mostly ignore the flow attributes like relative priorities, delay, jitter and so on.On the study of the datasheet of a switching chip of Dune Networks, this Thesis focuses on several aspects as below: improving the architecture of the MSM 3-stage Clos network, designing two scheduling algorithms: ACBS and CBSTDM, construction of a general model for simulation, and discussing the simulation results. Simulation results show that this improved architecture and scheduling algorithms can achieve high performance.First, this Thesis deals with the basic theories of switching technologies including switch fabrics and scheduling algorithms and looks forward to the future of this field. Then, the development of switch from single-stage to multi-stage is discussed; the differences on architecture and scheduling algorithm between single-stage and multi-stage are analyzed.Second, improvements have been made on the MSM 3-tage Clos, which can support distributed"pull"scheduling algorithms with better scalabilities.Third, on the base of the improved architecture, two scheduling algorithms have been designed: ACBS and CBSTDM, which both are distributed traffic scheduling schemes and totally different from the traditional fabric scheduling schemes. The traditional fabric scheduling algorithms push the data towards egress; however, ACBS and CBSTDM pull data from ingress. In ACBS, ingress and egress transfer asynchronous control messages to finish distributed"one-hop"scheduling. Simulation...