ADMISSION CONTROL AND RESOURCE-MANAGEMENT FOR MULTISERVICE ATM NETWORKS

We propose a resource allocation scheme that takes into account both t he QoS requirements and the characteristics of traffic streams. In our scheme, the decision to allocate buffer and bandwidth resources to a QoS class depends on whether buffering will be effective in reducing l oss probability for this class. We present analytical techniques based on the theory of large deviations for estimating the small loss proba bility. Furthermore, we address the problem of accepting calls with in complete traffic statistics. Existing proposals of traffic descriptors call for specification of only the first-order statistics such as the peak rate and the sustainable cell rate. To guarantee QoS, the admiss ion policy must be conservative by assuming the worst-case distributio n compliant to the traffic descriptor. We demonstrate that on-off sour ces exhibit such worst-case behavior. However, a static conservative p olicy results in under-utilization of network resources. We devise and analyze dynamic algorithms that improve the estimates of the actual b andwidth requirements for existing sources by collecting information o n-line. We show that a direct approach, where the quantity measured is the QoS of interest, combined with an extrapolation technique to spee d up the estimation process, offers a promising solution.