Prediction based resource allocation, a simulation experiment

We present an simulation experiment that can be used to asses the impact of using traffic prediction for online control on the performance of communication networks. A server with fixed output link bandwidth and impaired with high priority traffic is considered. The server shares its remaining available rate over K contending low-priority input streams queued at the input. An adaptive linear filter is used to provide an estimation of the future traffic load on each stream. The influence of the number of coefficients of the filter (window size) on the obtained predictions is discussed. The estimates for the future load on each connection are fed to a rate allocation algorithm that decides on how the available rate will be shared among the low-priority queues. Currently, the objective of this algorithm is to minimize packet loss by keeping all low-priority queues at the same size.