Survivable Computer Networks in the Presence of Partitioning

A small number of network components failures can cause high-speed networks (LANs) or System Area Networks (SANs) to be partitioned. This issue has been resolved with the introduction of redundant network components (ie. interfaces) into the network. A cost effective solution can be achieved with switch interconnects. Recently, such networks have been successfully implemented[6] with Myrinet networks. In this article, we present a "F cycle ring" (FCR) architecture which can withstand up to (F+1) failures without partitioning the switch interconnect. Our simple design is scalable to any ring size, given a lower limit on the number of switches. We focus on switches with a degree of four and computation nodes of degree two. We prove that (F+1) failures can be tolerated by the interconnect with several lemmas. Some examples of FCR are shown as an implementation of a full FCR network. Index Terms Partitioning, rings, meshes, torus, interconnection networks