On Long-Range Dependence in NSFNET Traffic

Abstract

Several recent studies indicate that (i) dependent packet arrivals can have a significant impact on buffer-overflow probabilities and delay characteristics at a network multiplexor, with long-range dependence showing a larger impact than short- range dependence, and (ii) that traffic in several different campus-level network environments appears to be long-range dependent. These two related issues are beginning to make a significant impact on traffic theory and network resource management. The objective of this paper is to show that long-range dependence extends to wide-area networks as well, and to identify the factors contributing to it. In particular, we find the presence of long-range dependence in packet level traffic over the NSFNET wide-area network. The analysis is based on raw traffic data collected from multiple NSFNET core switches distributed across the US. Investigation of the underlying processes for a subset of traffic created by TCP indicates that the (heavy-tailed) joint distribution of TCP conversation sizes and conversation transmission rates are the key factors contributing to long-range dependence. Since these distributions are primarily application-level characteristics and to a certain extent network independent, long-range dependence will be present in highly- multiplexed data networks that are running current TCP applications. The work of Steven M. Klivansky and Amarnath Mukherjee was supported in part by the National Science Foundation under grant NCR-9396299.