A Study Of Fair Dynamic Bandwidth Allocation Based On Burst Traffic Prediction In EPON

"Optical progress and copper retreat" has become a development trend of broadband access technology among telecommunication operators in market. EPON (Ethernet Passive Optical Network) systems have been considered as the most competitive solution to resolve "the first one mile" problem in all-optical networks because its long distance and wide available bandwidth. EPON combine the advantages of passive optical network and Ethernet, which can not only provide IP data service access seamlessly, but also expand its ability to multi-services support and QoS guarantee. EPON have a bright future which can support integrative access with multi-services such as data, voice and video.As a key technology, upstream bandwidth allocation is studied hot in EPON. With in-depth study on dynamic bandwidth allocation strategy and polling mechanism, a fair dynamic bandwidth allocation called BTP-DBA based on burst traffic prediction is proposed in the paper, the design goal of which is that the networks can achieve high bandwidth utilization and can support QoS fairly. Artificial neural networks are introduced in traffic prediction for the first time in EPON, and burst sorting polling is proposed in the algorithm simultaneously, including the strategy of bandwidth allocation based on the proportion of guaranteed bandwidth. The algorithm could be a feasible method for upstream bandwidth allocation, for which has a higher bandwidth utilization, less packet delay, and improves the QoS of lower priority service when in heavy traffic.The primary work and results are as follows:(1) The paper has designed a traffic prediction algorithm based on neural networks which are optimized by improved particle swarm optimization algorithm in EPON. The prediction algorithm is established by neural networks with three layers and one output. The neural networks are trained by PSO algorithm, and the PSO algorithm is improved by adding disturbance to particle with poor quality, which makes the prediction algorithm converge faster and more accurate. And the neural networks can adjust weight according to prediction error. As a result, the prediction is more accurate and adaptive.(2) The burst sorting polling mechanism is proposed in the paper. The OLT adjusts the polling order according to the burst traffic from each ONU during polling, and each ONU delays sending report message. As a result, the OLT can get richer and newer information about the queues of ONUs'. The new polling strategy can shorter the prediction range of the burst ONUs' queue, so it improves the accuracy of traffic prediction further. Besides, the OLT allocate bandwidth and grant time space more early. Thus the network can make full use of idle time, so the mechanism enhances efficiency of the upstream bandwidth.(3) A fair dynamic bandwidth allocation algorithm based on traffic prediction is proposed in the paper. It allocates bandwidth among ONUs and intra-ONUs according to the remedied bandwidth requirement after prediction and the proportion of guarantee bandwidth. The scheme not only reduces packet delay, but also guarantees fairness.(4) Finally, the EPON simulation platform is established based on OPNET, in which the BTP-DBA algorithm is verified. And the results show that the prediction algorithm based on neural networks optimized by improved PSO perform better in convergence and accuracy. The system adopted BTP-DBA gets higher bandwidth utilization and less packet delay. Furthermore, under the condition that the high priority service is ensured, the QoS of the low priority service is improved greatly when in heavy traffic.