Buffer Sizing results for RCP Congestion Control under Connection Arrivals and Departures

By: 
Ashvin Lakshmikantha, R. Srikant, Nandita Dukkipati, Nick McKeown, and Carolyn Beck
Appears in: 
CCR January 2009

Buffer sizing has received a lot of attention recently since it is becoming increasingly difficult to use large buffers in highspeed routers. Much of the prior work has concentrated on analyzing the amount of buffering required in core routers assuming that TCP carries all the data traffic. In this paper, we evaluate the amount of buffering required for RCP on a single congested link, while explicitly modeling flow arrivals and departures. Our theoretical analysis and simulations indicate that buffer sizes of about 10% of the bandwidth-delay product are sufficient for RCP to deliver good performance to end-users.

Public Review By: 
Darryl Veitch

This paper considers the effect of bottleneck buffer size in single-bottleneck networks using processor sharing, or more specifically, a max-min flavor of the RCP protocol, on a single congested link. It is one of few papers which consider the dynamics of flow arrivals and departures. The paper provides many interesting results on the problem of buffer sizing using RCP. Of particular interest are theorems 1 and 3, which provide very clean characterizations of the probability of buffer overflow, in the case of exclusively long flows, and exclusively short (non-reactive) flows, respectively. The overall conclusion is that, under RCP, buffers can be small, at no more than 10% of the overall bandwidth delay product, and simulation results at the end of the paper back this up for more realistic mixed traffic scenarios.
The paper is well organized and a pleasure to read. Although the analysis involves some fairly strong assumptions, it provides a useful view on the idea that communication protocols can provide good performance with buffer sizes that are very much lower than the bandwidth-delay product dimensioning rule in use today. A particularly nice result is an explanation for why the loss probability under a small buffer may actually decrease at higher load in some circumstances. To understand why, well you will have to read the paper.