Abstract
Scheduling algorithms for next generation multi Tbit/sec routers and switches are evaluated. Concurrently, available packet interconnections do not exceed a few hundreds Gbit/sec of aggregate throughput, making traffic modeling and computer simulations an extremely important tool. Here, traffic modeling is investigated in terms of arrival process and destination distribution. ‘Real world’ traffic models are proposed, and extensive simulations are carried out and compared to theoretical predictions. It is shown that conventional scheduling algorithms, such as round robin, perform very well under ‘convenient’ simulation assumptions, e.g., heavy load and uniform distribution assumptions. However, when traffic modeling is oriented towards modern applications such as bursty arrival process as in compressed video, or non‐uniform destination distribution as in Internet sites popularity, detailed and quantitative simulation results prove that algorithm performances are significantly degraded from over 95% to less than 60% of maximum throughput. A modified ‘adaptive’ round robin algorithm is proposed, yielding significant improvement of 20% in the router's maximum throughput.