Simulation-based results of weighted fair queuing algorithms for ATM networks

The behavior of the ideal General Processor Sharing (gps) discipline and di erent per-VC queuing algorithms approximating this ideal scheme, namely the Self Clocked Fair Queuing, the Packet by packet Generalized Processor Sharing, and the Virtual Starting Time disciplines, are studied in this paper via simulation. We speci cally consider a simple simulation con guration involving two Constant Bit rate (cbr) connections and several on/off connections (bursty tra c). This simple simulation experiment allows us to point out three important features of the gps and approximating disciplines. First, by adequately choosing the weight coe cients, these scheduling schemes can o er to cbr tra c almost Head of Line (hol) priority over on/off connections, to each of which, nevertheless, a minimum bandwidth is guaranteed. Second, gps and per-VC queuing disciplines, like the simple fifo scheme, is very sensitive to burst scale congestion phenomena. Finally, simulation results seem to indicate that the scheduling disciplines considered perform tra c shaping on on/off connections, which drastically reduces the burstiness of output tra c.