Re: How quickly does the queue size change at an Internet core router?

Hi, Leon,

At 05:45 p.m. 19-10-2004 -0700, you wrote:
>Hi Folks,
>
>Does anyone know how quickly the queue size changes at
>Internet core routers? Please give me a reference if
>possible.
>
>This question is very important to active queue
>management (AQM) algorithms. To see this, note that
>many AQM algorithms operate on the queue size, and the
>action taken by an Internet core router will take
>about an RTT to be perceived by a TCP sender.
>Therefore, if the queue size changes too fast compared
>to an RTT, AQM algorithms will simply not work.

The queue size RED uses to work with is not the instant queue size but some 
smoothed value calculated using EWMA.

Best regards,

Leonardo Balliache


>A more detailed description on this question is
>available at:
>http://www.cs.ucr.edu/~mart/204/CS_204_Project.html
>
>
>And here I quote the question:
>"Active queue management algorithms (like RED) try to
>prevent congestion by watching how the queue sizes
>fluctuate over time. If the queue size starts to get
>large and looks like it will soon grow even larger,
>the queue management algorithm attempts to cause the
>TCP senders who are using this link to reduce their
>respective transmission rates by randomly dropping a
>few packets before the buffer is completely full.
>However, the question is, can the queue management
>algorithm inform the TCP senders of impending buffer
>congestion quickly enough to be useful?  The typical
>round-trip delay for a TCP session that crosses the
>Internet is on the order of 10s or 100s milliseconds.
>Thus, if the queue management algorithm tells a TCP
>sender to slow down now (or, conversely, that it is OK
>to speed up again), the sender will be unable to react
>until one round-trip time later. If queue sizes change
>very slowly (i.e., the time between episodes of
>congestion is much larger than the round-trip delay),
>then the queue management algorithm can be quite
>effective. However, if queue sizes change very quickly
>in comparison to the round-trip delay then queue
>management will not work very well. "
>
>Thanks.
>
>- Leon
>
>__________________________________________________
>Do You Yahoo!?
>Tired of spam?  Yahoo! Mail has the best spam protection around
>http://mail.yahoo.com