Review: "Controlling High Bandwidth Flows at the Congested Router"

The paper proposes a congestion control mechanism called RED-PD (RED
with Preferential Dropping). The mechanism identifies high bandwidth
flows, which cause congestion at the router, using RED's packet drop
history and then monitors them to control their throughput by
probabilistically dropping packets. It only keeps partial flow state,
that is, state for high bandwidth flows only. Since these flows
constitute only a small fraction of the total bytes sent, controlling
their throughput results in a decrease in the drop rate at the output
and leads to higher throughput for other flows.
RED-PD consists of two parts:
1) Identifying high bandwidth flows: uses the drop history, which is a
reasonably random sample of incoming traffic and also represents flows
that have alrady been sent congestion signals, to determine which flows
to monitor. When the bandwidth of these flows is above the target
bandwidth, it must be controlled.
2) Controlling the bandwidth of monitored flows using preferential
dropping: The drop rate depends on the excess sending rate of the flow
and is done using a prefilter in front of the output queue. Unmonitored
flows are put directly at the output queue. The paper argues that the
mechanism provides fairness among monitored flows and protects
unmonitored flows from monitored ones, does not stave monitored flows,
does not protect monitored flows from general congestion at the link
because the output queue does not differentiate betwwen flows.
Furthermore, several things ensures stability and efficient utilization,
for example, if there is insufficient demand at the output queue packets
are not dropped, the dropping rate is adjusted over several intervals
(spaced out to allows the flow to respond) and both decrease and
increase of dropping probability is bounded in order to prevent
oscillations.
The paper presents evaluations for several aspects including
effectiveness of identification (improves with increasing sending rate,
an unidentified flow will be identified soon), fairness (can be
approxiamted by iteratively increasing and decreasing the pre-filter
dropping), response time to sudden changes in a flow's sending rate and
effect of the target RTT on identification of flows and bandwidth
received by monitored flows.
Finally, the paper talks about low state requirements and complexity of
their mechanism, and describes how they handle misbehaving flows by
declaring them as unresponsive and keeping them under tighter control.
However, they point out the possibility of false positives in this.
The paper's evaluation section is 'incomplete'. They refer to different
papers rather than presenting their results in a consice manner.
Furthermore, they simulate only a few conditions.
Received on Mon Oct 02 2006 - 11:28:07 EDT