Implications of longtailed traffic in networks: Analysis, measurements and management

Purdue University, West Lafayette IN

Investigators

Abstract

With the tremendous growth of the Internet and the ability to transport information at very high data rates has come the realization that traditional assumptions on the stochastic nature of network traffic ows are inadequate. A number of recent studies have demonstrated that long-range-dependence is an important characteristic of traffic ows in current networking infrastructures. On the other hand, the performance of networks and their ability to offer Quality of Service (QoS) depend on accurately capturing (and exploiting) probabilistically rare phenomena. Studies of network performance with long-range-dependent inputs are still very much in their infancy, and to-date have addressed only the computation stationary buffer over ow probabilities. In contrast, there is a crucial need to study the rare excursions of the underlying stochastic pro- cesses which capture the salient characteristics of network ows, and to understand how such excursions affect network dynamics. These issues are important from the point of view of transient characteristics of networks. For example the QoS perceived by users depends not only on the magnitude of loss but how it takes place. In some applications consecutive losses even if rare can make transmissions to be severely degraded while in TCP, losing a small number packets often can result in deterioration in performance. To study these effects we must develop new results on the duration of over ows and the amount of information lost during typical over ows. The goal of our research is to provide analytic insights, quantify performance, develop approaches to measure relevant traffic parameters and provide rules-of-thumb for use in resource management and scheduling traffic in networks with long-range dependent characteristics to meet QoS requirements. The specific tasks that will be pursued to achieve our objectives are: Study the buffer over ow asymptotics for multi-buffered systems with heterogeneous longtailed inputs. Develop results for the heterogeneous sub-exponential case. Characterize the transient behavior during over ow periods for systems with longtailed and sub-exponential inputs with a view to understanding actual QoS given to users. Develop methods for differentiating between traffic types and estimating relevant parameters. Incorporate the insights obtained from the above analyses into our overall objective to define engineering rules for resource management and scheduling various types of sessions in networks. As a result of these investigations we will have a complete catalog of results to handle the very diverse types of traffic which are present in networks. By considering both the stationary and transient characteristics we will provide insights on the dynamics of loss which is essential to provide fine tuning in order to deliver QoS at the application/user level which stationary analyses do not provide. The research spans fundamental modeling and analysis as well as statistical testing. On the theo- retical side the implications go beyond the traffic engineering context and are expected to contribute to applied probability, queueing analysis and asymptotic approximation. On the applications side, the proposed work will provide insights and results of potential use in next generation networks.