Controlling High-Bandwidth Flows at the Congested Router

Motivation: This paper proposes RED-PD to control high-bandwidth flows
at the congested router in which high-bandwidth flows are identified
based on the packet drop history and then are monitored to drop packets
and control their throughput.

Key Points:

1) The central idea of the paper is that only a small fraction of flows
are responsible for most of the bytes sent. A mere 1% of the flows
accounted for about 80% of the bytes and 64% of the packets. This shows
the skewedness of bandwidth distribution as well as justifies the
partial flow state approach.

2) RED-PD uses RED drop history to identify high bandwidth flows. RED
drop history is a reasonably random sample of incoming traffic and also
represents flows that have been indeed sent congestion signals.

3) The reason for using M lists instead of last K seconds seems not well
motivated in the paper. One reason could be that they want to use an
"average" to determine the high flows.

4) Preferential dropping is done using a pre-filter in front of the
output queue with a probability dependent on the excess sending rate of
the flow. Unmonitored traffic is directly put in the output queue.

5) In the evaluation, the authors have used fixed size packets (1000
bytes). This does not depict real world scenario and thus, raises some
suspicion on the results obtained.

6) The RTT R plays a significant role in the effectiveness of RED-PD and
it might be worthwhile in looking into the possibility of dynamically
varying R depending on the health of the network.

7) The protocol keeps a check on misbehaving flows by declaring them as
unresponsive and then penalizes them by tighter control although this
scheme is subject to false positives and negatives.

The results of RED-PD definitely look promising and are on obvious
improvement to RED.
Received on Sun Oct 01 2006 - 23:47:39 EDT