Long-range dependence in VBR video streams and ATM traffic engineering

Studies at Bellcore have shown that traffic rates in high-speed data networks exhibit ‘long-range dependence’ in their correlation, characterized by the so-called Hurst parameter H taking values in the range 0.5 < H < 1. This corresponds to a slower asymptotic decay of correlation than in finite-state Markovian processes, for which H = 0.5. While it has been shown that, in some situations, link-engineering based on Markovian traffic models could underestimate the bandwidth requirements for traffic with H > 0.5, there are other results showing the successful use of a Markovian traffic model to determine bandwidth requirements for variable-bit-rate video sources, which do have H > 0.5. ort here the results of an investigation undertaken to bridge the gap in our understanding of these two sets of results, and to examine the effect of the Hurst parameter on link-engineering for ATM traffic. Our investigation has shown that a curious ‘scaling’ property of streams with long-range dependence, derived in another recent Bellcore study, helps to bridge the gap by accounting for both sets of results. The analysis suggests that a high value of H is not, in itself, a reason to suppose that Markovian traffic models will lead to under-engineering of bandwidth on ATM links.