Summary: Controlling High Bandwidth Flows at Congested Router

Summary: Controlling High Bandwidth Flows

This paper presents a mechanism RED-Preferential Dropping - that provides
simplicity and protection by keeping flow state only for high bandwidth
flows. It uses RED packet drop history to detect high bandwidth flows and
preferentially drops packets from these flows.

The paper makes use of three important observations. First, controlling
high-bandwidth flows at the congested router significantly improves the
performance of the rest of the traffic. Second, packet drops from RED
management are unbiased sample of the incoming traffic as RED drops are
probabilistic. Third, high bandwidth flows in a given interval are a good
predictor of high bandwidth flows in the next interval.

RED-PD has two components: Identification of high bandwidth flows, and
Preferential dropping scheme. High-bandwidth flows identification makes
uses of RED drop history. Behavior of RED-PD remains the same as that of
RED as long as router has sufficient bandwidth to support all flows.
RED-PD partitions history into multiple(M) lists containing drops from
consecutive intervals of time. If it identifies flows with at least K
losses in M lists, then it's a high bandwidth flow. Preferential drop
component makes uses of the above observation to predict high bandwidth
flows. The pre-filter drop policy drops packets from each link based on
the flow specific probability. This preferential dropping not only
protects unmonitored traffic from monitored flows but also provides
fairness among monitored connections.

The rest of the paper evaluates the fairness and effectiveness of RED-PD
through simulations. The results show that identification probability
increases quickly as the sending rate of a flow increases. Also, the
speed of RED-PD's reaction depends on the ambient drop rate and arrival
rate of monitored flows i.e. probability of identification is higher if
either of them is higher. Final section of the paper provide little
insight into the state requirement complexity at the routers.

The simplicity of this solution makes it a viable alternative for Internet
Routers today because RED-PD doesn't require any changes to packet
headers. Also, it can be readily deployed for RED routers. The drawback of
the paper is that it does not present a good solution to identify
unresponsive flows.
Received on Tue Oct 03 2006 - 06:32:15 EDT