Performance Optimization of ATM Input Control using an Adaptive Leaky-Bucket Mechanism

This paper investigates a combined reactive/preventive congestion control mechanism. Emphasis is given to optimize the Leaky Bucket (LB) mechanism chosen for source traffic policing in ATM networks, examining performance characteristics such as reaction time, cell loss and cell delay. The performance of both buffered and unbuffered LBs is analyzed using the fluid-flow model. Closed form formulas are derived to calculate the mean queuing delay and mean reaction time of the LB. Numerical calculations and simulations are performed for different traffic sources which are characterized by their mean bit rate, peak bit rate and average number of bits generated during the burst. The results obtained confirm the performance studies done so far, showing that one LB is not sufficient to police all the source traffic parameters. It has been found that a three-fold trade-off exists between sensitivity to violation, reaction time and cell queuing delay. According to the analysis done a triple LB mechanism is proposed which provides an effective solution if tight control, fast reaction time and a small queuing delay are required. The LB parameters are then dynamically changed according to the delayed congestion feedback information received from the network. The preventive control policy is finally compared with the adaptive control scheme showing that even for large propagation delays, major performance improvements are possible by using an appropriate feedback policy.