Date of Award
Doctor of Philosophy in Electrical Engineering - (Ph.D.)
Electrical and Computer Engineering
Ali N. Akansu
H. Jonathan Chao
Internet growth and success not only open a primary route of information exchange for millions of people around the world, but also create unprecedented demand for core network capacity. Existing switches/routers, due to the bottleneck from either switch architecture or arbitration complexity, can reach a capacity on the order of gigabits per second, but few of them are scalable to large capacity of terabits per second.
In this dissertation, we propose three novel switch architectures with cooperated scheduling algorithms to design a terabit backbone switch/router which is able to deliver large capacity, multicasting, and high performance along with Quality of Service (QoS). Our switch designs benefit from unique features of modular switch architecture and distributed resource allocation scheme.
Switch I is a unique and modular design characterized by input and output link sharing. Link sharing resolves output contention and eliminates speedup requirement for central switch fabric. Hence, the switch architecture is scalable to any large size. We propose a distributed round robin (RR) scheduling algorithm which provides fairness and has very low arbitration complexity. Switch I can achieve good performance under uniform traffic. However, Switch I does not perform well for non-uniform traffic.
Switch II, as a modified switch design, employs link sharing as well as a token ring to pursue a solution to overcome the drawback of Switch 1. We propose a round robin prioritized link reservation (RR+POLR) algorithm which results in an improved performance especially under non-uniform traffic. However, RR+POLR algorithm is not flexible enough to adapt to the input traffic. In Switch II, the link reservation rate has a great impact on switch performance.
Finally, Switch III is proposed as an enhanced switch design using link sharing and dual round robin rings. Packet forwarding is based on link reservation. We propose a queue occupancy based dynamic link reservation (QOBDLR) algorithm which can adapt to the input traffic to provide a fast and fair link resource allocation. QOBDLR algorithm is a distributed resource allocation scheme in the sense that dynamic link reservation is carried out according to local available information. Arbitration complexity is very low. Compared to the output queued (OQ) switch which is known to offer the best performance under any traffic pattern, Switch III not only achieves performance as good as the OQ switch, but also overcomes speedup problem which seriously limits the OQ switch to be a scalable switch design. Hence, Switch III would be a good choice for high performance, scalable, large-capacity core switches.
Chen, Feihong, "Design and analysis of a scalable terabit multicast packet switch : architecture and scheduling algorithms" (2000). Dissertations. 395.