Motivated by the need to provide quality of service (QOS), this dissertation analyzes loss performance and, specifically, the temporal correlation of cell losses in ATM switches with bursty traffic. While previous investigations of ATM loss performance have focused mainly on the evaluation of cell loss probability and conditional loss probability, the emphasis here is placed on a much more refined measure, namely the probability distribution of a cell-loss period. To accurately assess the impact of traffic characteristics on such a measure and QOS provision, a new definition of traffic burstiness of a source is proposed in terms of a non-bursty source. This definition captures both the first-order and second-order burstiness of a source and exhibits the special features that existing definitions do not provide. To model the ATM slot-based transfer technique, the evaluation of loss correlation is based on a discrete-time, finite-state Markov model, representing both the switch and its traffic, where source traffic at each input port is represented by an $(M+1)$-state source model (M ON states and one OFF state). In term of this underlying system model, various system-oriented and user-oriented loss measures are formulated. In addition, for purposes of application, this dissertation further develops a simple but accurate approximation of the tail distribution of a loss period, together with analyses of the insensitivity of this distribution to model parameter values. In the dissertation, the direct analyses of the system model are integrated with high-level stochastic activity network (SAN)-based analyses to determine the influence on the distribution of a loss period and to demonstrate the robustness of the approximation of the distribution. In particular, numerical studies of various loss measures and comparisons of the exact and approximate formulations are obtained by using SANs to specify corresponding instances of the system model, thus permitting automatic model construction and solution with a SAN-based ...