From: T Scott Saponas (ssaponas@cs.washington.edu)
Date: Wed Oct 20 2004 - 06:21:55 PDT
Review by T. Scott Saponas
In "Congestion Control for High Bandwidth-Delay Product Networks" a new Congestion Control method is proposed that outperforms current methods for high bandwidth-delay product networks. The authors point out that current methods do not work on these networks because when the delay is large and dropped packets are used for signaling congestion one has to be aggressive about backing off and conservative about ramping up. This causes wild oscillations because back offs are so extreme (slow start to ½ of window then only increase window by one packet per RTT). Because these are high bandwidth networks the additive increase is so slow that much of the bandwidth is not utilized for large periods of time. The authors proposes using a congestion header in each packet that informs the routers what each host thinks the current congestion is and lets the bottleneck router feedback by overwriting the value what the current congestion actually is. This method has the advantage of being able to tell flows exactly how much to back off so that full bandwidth can still be used even when there is congestion. It also has the advantage of being able to control fairness correctly.
The authors proposed solution has a couple of elegant features. They have each router estimate the average RTT by looking at the RTT field in the congestion header of each packet. Then they only give congestion feedback in at least average congestion window intervals. This is neat because it allows the router to see the effect of its last decision before making another decision. Another interesting part of the proposed solution is the abstraction of an Efficiency Controller (EC) and a Fairness Controller (FC). This allows the EC part of the router to just calculate total traffic change required and not worry about fairness and the FC part just figure out how to split the current bandwidth among flows. Thus, if either policy needs to change only one of these parts of the router needs to be modified.
The downside to this solution is it puts more data into the headers further reducing the efficiency of each packet. It also adds a little bit more complexity into the routers in the form of some more calculations per packet. However, these downsides appear to be reasonable tradeoffs for being able to utilize the fill bandwidth of high bandwidth large delay lines.
This paper is quite relevant to today's internet since higher bandwidth connections are always around the corner and this solution allows us to utilize that bandwidth even though the latency doesn't get lowered. The authors do a good job of making this solution particularly relevant by showing that it can exist with TCP and wouldn't add much complexity to the router.
This archive was generated by hypermail 2.1.6 : Wed Oct 20 2004 - 06:21:59 PDT