Controlling High-Bandwidth Flows at the Congested Router

Paper Review: Controlling High-Bandwidth Flows at the Congested Router

The paper introduces RED-PD, a method for controlling congestion at
the router. RED-PD is RED (random early detection) with Preferential
Dropping. In RED-PD, as opposed to RED, state information is only
maintained for high-bandwidth flows. The reason for this is that most
traffic is considered to be ?web mice? (short HTTP flows) and
maintaining state information for these flows is wasteful. Although
RED-PD would be unable to provide the same level of fairness as RED,
the paper argues that a high level of fairness is not required for
best-effort traffic. RED-PD attempts to limit high-bandwidth flows to
a target bandwidth by dropping packets at its output queue. RED-PD is
only active when the router is in a congested state, identified by
monitoring average queue lengths. Essentially, RED-PD aims to improve
performance for low-bandwidth flows using minimal state information
making use of the fact that at certain times 1% of flows can account
for 80% of the bytes in a router queue.

RED-PD has two parts. First it must identify those flows which are
consuming the most bandwidth and then in turn control the bandwidth
provided to these flows. RED-PD makes use of a sliding history of the
RED packet drop history to identify high-bandwidth flows. Traffic
that is being monitored (high-bandwidth) is subject to a probabilistic
drop, when in a congested stated, at the output queue pre-filter. The
traffic that is not being monitored is put in the queue directly.

In simulations, it was shown that as a flow?s sending rate was
increased, the probability of being identified as a high-bandwidth
flow increased and thus increased its chances of being throttled. A
fairness simulation showed that RED-PD is able to protect
low-bandwidth flows and control high-bandwidth ones. The next
simulation showed the response time to a flow increasing its sending
rate. The flow was subject to packet drops very quickly and its
sending rate was eventually controlled. Although the results are not
clear, it was shown that the time to bring down a high-bandwidth flow
decreases as the flow uses more bandwidth (increasing the overall drop
rate). This property shows that the more congestion a flow causes,
the faster it will be controlled. Next, a quick calculation showed
that the memory requirement for the state information is very
reasonable and does not require fast memory. Memory requirement was a
concern in the fair queueing scheme, an issue this paper has
addressed. A question which still remains is whether this scheme is
economically feasible. One advantage in its favour is that it is
partially deployable making it a realistic possibility. The
processing requirement was also shown to be minimal making it
worthwhile to implement a router utilizing RED-PD and running more
advanced real-world tests.

As a final note, the paper was well-written and well-organized. The
author?s seem to have a strong understanding of the related work which
strengthened the argument presented in this paper. The paper was
clear and concise. The use of graphs of charts complemented the
analysis section effectively.

-- Nadeem Abji
Received on Mon Oct 02 2006 - 16:51:08 EDT