Controlling High-Bandwidth Flows at the Congested Routers

   
  Controlling High-Bandwidth Flows at the Congested Routers
   
  This paper presents RED-PD for congestion control in routers. This mechanism uses the RED packet drop history routers to detect high-bandwidth routers in times of congestion, and preferentially drops packets from these flows. RED-PD combines the simplicity of FIFO with the protection of full max-min fair techniques. RED-PD is based on the fact that a small fraction of flows are responsible for most of the bytes sent over the network.
   
  There are two components in RED-PD: identifying high-bandwidth flows and controlling the bandwidth obtained by these flows. Since RED drops are probabilistic, and not the result of a buffer flow, they can be considered as reasonably random samples of the incoming traffic. RED-PD keeps this history in M lists and identifies flows with lost pockets in more than K lists as high-bandwidth flows and monitors them. Packets from monitored flows are dropped in the pre-filter with a probability dependent on the excess sending rate of the flow. This mechanism protects unmonitored traffic form the monitored flows, provides relative fairness among monitored flows, does not starve monitored flows and does not protect the monitored flows from the general congestion at the link.
   
  Different aspects of RED-PD are evaluated through simulation in the paper. These aspects are probability of identification, fairness, response time, and effect of the target RTT. Simulation results show that identification probability increase rapidly as sending rate of a flow is increased, and if a high-bandwidth flow escapes identification from a particular round, it will be identified soon. From fairness point of view, while in RED each flow receives a bandwidth share proportional to its sending rate, with RED-PD all the flows receive roughly their share. The speed of RED-PD’s reaction depends on the ambient drop rate and the arrival rate of the monitored flow. This leads to the result that if a flow increases its flow to a very high level, it will be brought down quickly. The final aspect of RED-PD is its simplicity and its low need in memory. It maintains very little state, and its fast-path operations are also the simplest.
   
  The paper is well organized and clear. It has explained the background information and different previous approaches to the problem well, and various details are developed well.
   
   
   

                                 
---------------------------------
Want to be your own boss? Learn how on Yahoo! Small Business.
Received on Mon Oct 02 2006 - 21:10:05 EDT