A neuro-forecastwater-filling scheme of server scheduling

Abstract

Dynamic server scheduling schemes in queuing systems accommodating delay-sensitive traffic need to address the tradeoff between efficiency and fairness. For delay-sensitive traffic, threshold-exceeding delay or equivalently loss is used as a measure of efficiency. In this paper, a pair of dynamic server scheduling schemes for queuing systems accommodating delaysensitive traffic are compared. Each scheme consists of two components. The first component attempts at forecasting the arriving traffic patterns of the sources sharing the server bandwidth and the second component makes the assignment of server bandwidth among the sources. The schemes utilize BFGS and resilient backpropagation learning in perceptron neural networks to forecast the arriving traffic patterns, respectively. Once the traffic patterns are forecast, the schemes rely on water-filling to make the server bandwidth assignments max-min fair. Our simulations reveal the efficiency and fairness characteristics of the schemes.