Understanding Buffer Management for Cut-through 1D Rings Cruz Izu 1 , Ramon Beivide 2 1 School ofComputer Science, The University of Adelaide, Adelaide SA 5001, Australia, cruz@cs.adelaide.edu.au 2 University of Cantabria, Avda Los Castros s/n, 39005 Cantabria, Spain mon@atc.unican.es Abstract. This paper describes the impact that buffer management has on net- work performance for a cut-through 1D ring. Such network provides only one routing alternative between each pair of nodes. The network is kept simple to il- lustrate the significance of buffer management for all virtual cut-through net- works. Besides, 2D and 3D torus, among other networks, can be seen as a col- lection of interlinked 1D rings. The simulation results will show that the key to maximum sustained throughput is to reserve half of our buffer capacity for tran- sit packets. This increases link utilization by maximising packets at the sending node and empty buffers at the receiving node. 1 Introduction Ring networks have been widely used to implement local, campus and metropolitan area networks. Extensions of 1-D rings such as Chordal Rings of degree three and four have been considered in the literature [2]. Popular extensions of the ring are the 2-D and 3-D Torus, frequently used to interconnect highly-coupled parallel systems [11]. An ideal router has infinite buffer space but will only buffer a packet when the output ports to which the packet is routed are busy. In real life, buffers are finite and have finite bandwidth. Buffer capacity in the order of a few packets per input link is suffi- cient to cope with the traffic variations of a non-saturated network [5]. Further buffer capacity has little impact on network response. Most routers use input FIFO queues because of their simple implementation that translates into low access time and larger buffer capacity [6]. On the other hand, buff- ers may be better used in centralized or output queues, with the additional advantage of eliminating the head-of-line blocking (HOB) exhibited by input FIFO queues. Full link utilization is achieved if we always have a packet requesting the link and if the flow control mechanism allows that packet to use the link. Most of the network research has focused on issuing the maximum number of requests from their incom- ing packets, by finding organizations that eliminate HOB within a reasonable cost in