Flow Control for ABR Service in ATM Networks Using Modular Neural Networks

For the case of HME-based controllers, we see that queue length increases to a large peak at the initial stage, followed by a transient period of less than optimum link utilisation. This corresponds to the period before HME has learnt, because we h a ve used random initial weights for the HME. In actual applications, the combination of both on-line and o-line training may be necessary. The initial weights of HME should be set to near optimal values, which could be obtained by o-line training using simple simulation models. Another way to reduce this transient eect is to force the controller to look into actual queue length during the initial stage, i.e. if actual buuer queue length is greater than the threshold value Q t , actions for congestion have t o be taken even if the HME s output indicates non-congestion. 10 Conclusion In this paper, we h a ve proposed a HME-based intelligent congestion control scheme, which is a direct enhancement o f t h e s c heme proposed by S i u & Tzeng 1. I t i s s h o wn to be superior in keeping queue length low, while maintaining high throughput in an ATM network. For the case where distances between sources and the switch are large, the ability to predict congestion becomes increasingly important, as it wo u l d b e t o o l a t e t o t a k e corrective a c-tions after congestion has already occurred. The introduction of a HME-based controller into an ATM switch increases the complexity of the switch, but it may b e w orthwhile, considering the beneets it can bring. The experiments described in this paper have assumed persistent sources, i.e. all sources will attempt to transmit cells at their highest possible speed 1. In practice, data traac is highly bursty. Our future work will investigate the performance of the HME-based scheme with such traac. Table 3: Throughput and average queue length at the end of 200ms of simulation time. Congestion Control Scheme Total number of cells Average received at destination queue length Link set 1 without HME (S&Z1) 67767 3.9103 Link set 1 with HME (HME1) 67017 2.1033 Link set 2 without HME (S&Z2) 60481 8.8762 Link set 2 with HME (HME2) 63465 3.9153 Table 3, it is observed that the HME-based controller has an average queue length …