RCBR: A Simple and Efficient Service for Multiple Time-Scale Traffic

摘要

Compressed video traffic is expected to be a significant component of the traffic mix in integrated services networks. This traffic is hard to manage, since it has strict delay and loss requirements, but at the same time, exhibits burstiness at multiple time-scales. In this paper, we observe that slow time-scale variations can cause sustained peaks in the source rate, substantially degrading performance. We use large deviation theory to study this problem and to motivate the design of Renegotiated Constant Bit Rate Service (RCBR), that adds renegotiation and buffer monitoring to traditional CBR service. We argue the the load placed on signalling by RCBR can be handled by current technology. We present a) an algorithm to compute the optimal renegotiation schedule for stored (off-line) traffic, and b) a heuristic to approximate the optimal schedule for online traffic. Simulation experiments show that RCBR is able to extract almost all of the statistical multiplexing gain available by exploiting slow time-scale variations in traffic. In more general terms, we believe that a clean system design must match control time-scales to the time scales over which the workload varies. RCBR works well because it makes intelligent use of this time-scale separation.