Previous work has demonstrated that long-range dependence (LRD) and multiple time-scales are important characteristics of realistic traffic. LRD traffic that exhibits multiple time scales poses a significant challenge to performance guarantee. We examine the impact of LRD in traffic, and address issues concerning how to compute and guarantee loss performance for LRD multiple time-scale traffic under realistic networking conditions.

Conventional analysis techniques of queueing systems typically assume that the stochastic process under study is already in steady state. This assumption is, however, not valid if the life cycle of the process is not large enough. Previous work in transient analysis of queueing systems usually focuses on Markov models. This paper, in contrast, presents an analysis technique for investigating transient performance of queueing systems that are not necessarily Markovian.

Performance-oriented studies typically rely on the assumption that the stochastic process modeling the phenomenon of interest is already in steady state. This assumption is, however, not valid if the life cycle of the phenomenon under study is not large enough, since usually a stochastic process cannot reach steady state unless time evolves towards infinity. Therefore, it is important to address performance issues in transient state.Previous work in transient analysis of queueing systems usually focuses on Markov models. This paper, in contrast, presents an analysis of transient loss performance for a class of finite buffer queueing systems that are not necessarily Markovian. We obtain closed-form transient loss performance measures. Based on the loss measures, we compare transient loss performance against steady-state loss performance and examine how different assumptions on the arrival process will affect transient loss behavior of the queueing system. We also discuss how to guarantee transient loss performance. The analysis is illustrated with numerical results.

Although it is evident that traffic correlations may have significant performance impacts on networks, it is still not clear how correlated traffic may deteriorate network performance. This paper demonstrates the catastrophic nature of network performance degradation induced by traffic correlations, that is, as long as the correlation time is large enough, bistability may appear in network behavior, which may cause the network to change abruptly from one stable state to another. This catastrophic behavior may have significant impacts on traffic control and engineering.

We analyse the impacts of the long-term correlations discovered in data and video traffic on congestion control in ATM-based local area networks based on catastrophe theory. Our work shows that long-term correlations may cause the network performance to degrade catastrophically even if the cell rate of the traffic source can be adjusted dynamically according to the network load and the peak cell rate can be bounded effectively. Experimental evidence exists supporting our analytical results. Therefore, without taking traffic correlations into account, congestion control and quality-of-service guarantees may not be effective.