Direct interconnection network has been widely used in supercomputer systems. Recently, it is considered to be used to build terabit router by the industry. This paper presents a distributed and scalable switching fabric based on a new direct interconnection network. It is a scalable topology and can be expanded in two ways easily, thus minimizing the initial investment of service providers. Its distributed control can offer low hardware complexity. Virtual cut through switching is used to achieve high throughput and low latency. The quality of service is guaranteed by introducing virtual channels based on the concept of DiffServ. The fault tolerant and load-balanced routing algorithm can offer deadlock and livelock freedom. It helps the network continue to work even with faulty parts existed. Finally, the performance of the proposed switching fabric is evaluated under various conditions. The results show it can outperform its counterpart in latency and throughput. It achieves terabit throughput with average latency of 1 us or so.
展开▼
