Dynamic resource management considering the real behavior of aggregate traffic

This paper explains why the theory of effective bandwidth is in general not applicable to characterize realistic traffic sources. Moreover, we show that a static allocation of network resources (bandwidth and/or buffer space) based on large deviation theory can be highly inefficient when the real statistical behavior of traffic is taken into account. As an alternative, we propose a dynamic resource management scheme based on prediction techniques. As a specific example, we apply this scheme to a Differentiated Service (DiffServ) Internet environment, in which the goal, besides policing the incoming traffic, is to optimize the use of network resources, thus minimizing the probability of occurrence of violations of contract guarantees. The performance of the proposed scheme is evaluated via simulations and our results show the superior performance of the new algorithm in terms of buffer overflows, output link utilization, and jitter, as compared to currently used policing and shaping mechanisms.