Paper: Controlling High-Bandwidth Flows at the Congested Router
Name: Robert Danek.
Course: CS2209, Fall '06.
This paper discusses the problem with traditional FIFO
(First-In-First-Out) congestion control schemes in routers, and presents
a new mechanism, called Random Early Detection with Preferential
Dropping (RED-PD), for solving the problem.
FIFO schemes for congestion control are not fair in that
high-bandwidth flows through a router can consume an excessive amount of
bandwidth while starving other flows. Existing alternatives to this
scheme require maintaining per-flow state for all flows going through a
router. This, however, is overkill, since most flows are short-lived
(such as flows for HTTP requests/responses).
The paper's scheme, RED-PD, maintains state information for a select
number of high-bandwidth flows. Identifying the high-bandwidth flows is
done by monitoring the packet drop history from RED. Once a flow has
been identified as high-bandwidth, any incoming packets from that flow
go through a "pre-filtering" phase where the packet is dropped with a
certain probability. The goal of the pre-filtering phase is to reduce
the bandwidth used by the high-bandwidth flow to a more reasonable
target bandwidth. In doing so, more bandwidth becomes available for
other flows.
The authors run a number of simulations using network simulator
(ns), examining the effectiveness of the scheme based on a number of
different factors. These factors include the probability of a flow being
identified as high-bandwidth, the fairness of allocation of bandwidth
amongst flows, and response time to rapid changes in the sending rate of
a flow.
Overall this was a good paper. The authors presented the details of
their scheme concisely, and then were able to demonstrate its
effectiveness via a number of simulations.
Received on Mon Oct 02 2006 - 15:58:43 EDT
This archive was generated by hypermail 2.2.0 : Mon Oct 02 2006 - 16:51:11 EDT