Contribuciones en arquitectura de redes de conmutadores transparentes Ethernet de altas prestaciones

Abstract

At present, campus and data center networks need high performance and manageability, within a reasonable price, especially for data center networks in which the tendency is to use more commodity devices instead of more complex and expensive alternatives. Ethernet switched networks turn up to be the first option in this field. Nevertheless, the adoption of the Spanning Tree Protocol (STP) as the paradigm to propagate messages through the network and avoid loops is a manifest limitation of the performance and size of Ethernet networks. While the two main recent standards, Shortest Path Bridging (SPB) and Routing Bridges (RBridges, TRILL) allow utilization of all infrastructure links to obtain shortest paths, but these standards use a link-state routing protocol (IS-IS), operation at layer two, and fail to follow the inherent simplicity of pure layer two transparent bridges. This thesis presents several contributions on Ethernet transparent bridges from the group generically denominated shortest path bridges to solve the big limitations that the spanning tree impose, but characterized to avoid the use of any routing algorithm, main approach followed by standards and proprietary proposals. In this way, the complexity of using a link-state routing protocol (like IS-IS in SPB and TRILL) is avoided, while taking advantage of the broadcasting-to-learn basic conception of pure transparent bridges in order to explore all paths in the network by using all the possible links of the topology. With this principle, we introduce the All-Path family, a family of Ethernet transparent bridging protocols that offers minimum latency paths with automatic load balancing, which can be adjusted to different topology requirements in scalability and load balancing. We also present Torii-HLMAC, a distributed, faul-tolerant, zero configuration protocol with multiple tree-based addressing and forwarding and on-the-fly path repair, specific for data center networks, primarily the so-called fat trees.