| One of the main advantages of the Asynchronous Transfer Mode (ATM) Broadband-Integrated Services Digital Network (B-ISDN) is the efficient sharing of the network resources through statistical multiplexing of variable-rate traffic streams. Although buffering is provided at the network nodes to relief traffic contention and absorb traffic fluctuations, the burstiness of the ATM traffic presents challenges to guarantee the multiple Quality of Service (QoS) requirements. The effective bandwidth provides an estimation on the bandwidth reservation and admission control, however, its overestimation and underestimation occurs. The ATM traffic bandwidth prediction seems to be an effective way for ensuring the ATM service QoS. During the recent years, neural networks have been utilized as the predictive mechanisms for the ATM traffic bandwidth predictors. However, limitation exists in the current methods, the neural network parameters need to be specified based on the traffic spectrum, or the traffic details have to be well known before those methods could be employed. During our investigation, a novel mechanism is proposed. The Wavelet decomposition techniques are used for extracting the bandwidth components in ATM traffic. As a result, the Effective Dynamic Bandwidth is introduced. A recurrent neural network with adaptive learning rate is then used for predicting the Effective Dynamic Bandwidth. The algorithms are verified using the Variable Bit Rate (VBR) MPEG traffic volume trace data. The algorithm is also effective on Available Bit Rate (ABR) LAN traffic. Discussion are extended to the future researches in the realtime implementation and the implication with multiple ATM nodes. |
