Review - RED-PD

Review of the paper "Controlling High Bandwidth Flows at the Congested
                    Router"
by Ivan Hernández

The paper presents a congestion avoidance (CA) mechanism based on
Random Early Detection (RED). The new CA is called RED-PD, because it
adds to the original a Preferential Dropping mechanism. The
preferential dropping is directed to high bandwidth flows, thus the
new solution favors flows with small values.

First, RED-PD has to identify the high bandwidth flows, so it uses the
the RED drop history -- remember that RED will drop packets of a flow
with a probability proportional to the flow's throughput through the
router --. RED-PD partitions in M lists. If RED-PD identifies flows
with a given number of losses in M lists, RED-PD will monitor these
flows.

Now, each time a packet arrives to the router , the router checks if
it is of a monitored flow. All packets from monitored flows are
dropped with a probability proportional to the excess sending rate of
the flow, unless that there is insufficient demand at the output
queue. Unmonitored packets are send to the RED process -- which will
keep generating new drops that will use the identifying process -- and
then to the output queue. If a monitored flow reduce its sending rate
enough, RED-PD will reduce the dropping probability to a negligible
value, and then the flow will not be monitored.

RED-P does fine in the simulations, it achieves fairnes between flows
with disparate bandwidths, and also shows how it protects
low-bandwidth flows. This protection is thanks to the monitoring and
preferential dropping of the high-bandwidth flows.

This paper gives an elegant CA mechanism, that does not need to keep
track of all the flows through the gateway, and thus it will need
little memory on the gateway. It is interesting to see how the authors
improved a good previous solution (RED) and how at the end they get to
a new and interesting approach.
Received on Mon Oct 02 2006 - 20:57:06 EDT