Congestion avoidance in high-speed interconnection systems

Abstract

Abstract In this paper we study the problem of scheduling packets through a network interconnecting several slow devices. Such a system might be high speed local area network (HSLAN) or geographically compact metropolitan area network (MAN) interconnecting several low speed LANs. The system under study is modeled by several input queues feeding data through the interconnection network (IN) to an output queue. We concern ourselves with the avoidance or minimization of overflow at the inputs and output. We study the behavior of four congestion avoidance policies that differ from each other according to the type of information passed between the queues. The most complex protocols use queue length information whereas the simpler protocols use either no information or packet age information. We compare the performance of these policies through a combination of analysis and simulation under the assumption of negligible delays through the IN. We observe that the more detailed the information, the better the performance. Howeve, if most of the buffers are allocated to the output queue, then there is little difference between these policies. This suggests that simple protocols may work well under such an allocation. We also study the issue of fairness when the arrival processes to the input queues are not identical. We observe that the policies based on queue length information provide fairer treatment when the performance metric is probability of loss and that simple policies provide fairer treatment in the case of mean packet delay. Last, we discuss the potential impact of our results to a system in which transit delays within the IN are not negligible.