Comparison of buffer allocation schemes in ATM switches: complete sharing, partial sharing, and dedicated allocation

The paper presents a performance comparison of three different buffer allocation schemes in ATM switches under bursty traffic conditions. It is well known that complete buffer sharing offers the highest throughput and lowest loss rates when traffic is not bursty. This is due to the statistical gains that are obtained from sharing buffers among outputs. One might expect that these statistical gains would only increase under bursty traffic conditions, but this is not always the case. Completely shared buffers may be plagued by congestion arising from a few bursty calls unfairly occupying the entire buffer space to the exclusion of other calls. Thus, under bursty traffic conditions there are two competing forces in a completely shared system: gains due to sharing buffers and losses due to unfairness. This trade-off becomes very important within the context of the bursty ATM environment. Dedicated queues are intrinsically fair, but they do not enjoy the statistical gains of sharing. Partial buffer sharing may offer a compromise between complete sharing and dedicated allocation by obtaining some gains from sharing while maintaining a degree of fairness. The multinet switch is one example of a partially shared system.<
>