Buffering packetised data: an examination of the crossover effect

Buffer analysis has traditionally relied heavily on the use of traffic models with short range (exponentially decaying) autocorrelation, e.g. Poisson and Markov modulated Poisson processes. Recent literature has suggested the presence of traffic with slowly (hyperbolically) decaying autocorrelation functions; these sources collectively and individually seem to be exhibiting the phenomenon of long range dependence (LRD). Furthermore, it has been claimed that this sort of traffic may exhibit behaviour, when buffered, that is advantageous in that a smaller buffer may be required than would be the case with non-LRD traffic. This phenomenon has been named the crossover effect, and the authors quantify its relevance to ATM networking, using simulation studies of a homogenous mix of chaotic sources. These show that there is a crossover effect, but the key finding is that this effect is much nearer to a buffer length of zero than was predicted by previous theoretical studies, actually occurring at a value so low that it has no practical significance