On CBR Service

Matthias Grossglauser, Srinivasan Keshav

Abstract: We investigate the performance of CBR traffic in the context of large-scale networks, where many connections and switches coexist and interact. We develop a framework for simulating such networks, decoupling the influence of breadth and depth. Our results are briefly as follows: we found that a Poisson stream is a good approximation to a superposition of many CBR streams with differing phases and bandwidths. Delays incurred by a reference stream with cross traffic composed of many CBR streams with different bandwidths and phases do not exceed a few cell times even under heavy load, which means that buildout buffers of 10 to 20 cells seem to be sufficient after traversing 20 switches. CBR traffic can be efficiently served by the first come first served (FCFS) scheduling discipline, which has the least implementation cost. Surprisingly, the round robin (RR) and weighted round robin (WRR) disciplines perform worse than the FCFS, despite their greater implementation complexity. We also compare an analytical approximation method based on the multiclass parametric decomposition method, with the simulation results and found it to be suitable for estimating the end-to-end delays for the FCFS discipline.