A Fast Switched Backplane for a Gigabit Switched Router

This paper presents a strong case for routers with a switched backplane. Initially, the authors provide a very useful overview of router architecture, breaking the path of a packet down into the forwarding decision, the backplane and the output-link scheduler. Router architectures are moving away from the use of shared backplanes as they can easily become congested and limit the performance of the system. Using crossbar switches instead allows multiple packets to be simultaneously transferred across the backplane, thereby increasing the aggregate bandwidth of the system. Simply making the backplanes faster has proven to be insufficient to keep up with common high packet arrival rates, so switched backplanes are the optimal solution. The authors briefly discuss why handling variable-length packets leads to lower throughput - essentially, abnormally long packets at the input can hold up other packets instead of being forwarded to a starved output. HOL blocking severely limits the throughput since the cell at the head of the FIFO input queue blocks cells behind it from access to the outputs. The authors present a solution in virtual output queueing (VOQ), which was also discussed in the previous paper. Next, the iSLIP algorithm is introduced as an input-output matching algorithm that provides a fast, high-throughput, starvation-free solution.

I truly enjoyed reading this paper. First of all, the background material was reviewed very well before jumping into any algorithms or proposals. Router architecture was covered in great detail and the decision to focus or not focus on each aspect of this architecture was explained quite well. This paper should definitely be part of the syllabus. One thing I wasn't sure about was whether these ideas were already being implemented at the time or whether they were proposed ideas for future implementation.