Effective strategy to compute forwarding tables for infiniBand networks

InfiniBand is very likely to become the facto standard for communication between processing nodes and I/O devices as well as for interprocessor communication. The InifiniBand Architecture (IBA) defines a switch-based network with point-to-point links that support any topology defined by the user. Routing in IBA is distributed based on forwarding tables, and only considers the packet destination ID for routing within subnets. Up*/down* routing is the simplest and most popular routing algorithm for irregular topologies. Unfortunately, up*/down* routing cannot be used in IBA switches because it may leads to deadlock. In this paper we address this issue, proposing an easy-to-implement strategy to complete up*/down* forwarding tables for IBA switches that guarantees deadlock freedom, and is effective whatever the methodology applied to compute up*/down* routing tables. Preliminary evaluation results modeling an InfiniBand network at register transfer level show that the proposed strategy allows up*/down* routing algorithms to be implemented on InfiniBand networks with minimal performance degradation.