DESIGN OF EFFECTIVE BUFFER CONTROL POLICIES FOR ATM NETWORKS

Due to limited buffers and highly unpredictable traffic loads, the des ign of effective buffer management policies that minimize cell loss is critical to ATM networks. We outline the important characteristics of optimal buffer control policies for a single node, and using a dynami c programming formulation for a simple two-node network, gain insights into the much harder general multi-node problem. The key result we pr esent is that simple non-work conserving policies which use limited fe edback from adjacent nodes can reduce cell loss significantly. We cons truct a feedback-based control that uses simple ''stop-and-go'' rules. This control is then enhanced to react to rapidly changing network co nditions. This new control is called the Adaptive HILO policy. Simulat ion studies of complex networks with large delay-bandwidth product sho w that the Adaptive HILO policy performs impressively under diverse co nditions. Using intuitive arguments, we demonstrate that this policy e ffectively utilizes unused buffer capacity in upstream buffers during periods of heavy load in downstream buffers.