Goings on at bufferbloat.net…

Kathie Nichols wrote in pointing out an insightful talk of Van Jacobson’s entitled A Rant on Queues. She also points out:

Also, as someone who spent a lot of time examining the dysfunctionality of 93 RED and various changes Van came up with in response to the problems: examine any AQM very carefully. It’s likely not working the way you think it is. What I learned is that almost anything works with nicely behaved long-lived TCPs and that almost nothing works well with mixtures of mice and elephants. It’s also instructive to examine what is really happening when your AQM drops packets. We found that 93 RED spent a lot of time doing what we called “forced drops” which means you are in the “drop everything” part of your control law. What we would find with 93 RED were a bunch of back to back drops. There’s tons of stuff on this in our unfinished “RED in a different light” paper, but look at the first 10 slides or so of

“We have Met the Enemy and [S/]He is Us”:A View of Internet Research and Analysis and you’ll see some of it without wading through a lot of boring prose. So, please, please, don’t make a “fix” that is perhaps worse than the original problem.

I think we need to take Kathie’s first hand experience to heart.

Richard Scheffenegger has done a quick NS2 simulation/animation of bufferbloat well worth a quick look.  Helping him out on that would be a great service.

Dave Täht came up with an interesting idea of possibly using NTP data to get a view into bufferbloat on a global scale. In short, bloat badly disturbs RTT’s and that may be a way to see what’s going on. He’s calling it the “cosmic background bufferbloat detector“.

About these ads

2 Responses to “Goings on at bufferbloat.net…”

  1. Ivan Barrera Says:

    93 RED was very prone to instabilities. No wonder such algorithm would fluctuate between the two highly non-linear parts of the piece-wise linear function describing RED. Moreover, the actual coefficient of the IIR Filter (EWMA), was empirically set to 0.02. Which in future versions was linked to the link capacity, and actually, based on my research could also be linked to an estimate of the number of flows to improve performance.

    RED if something, was a pioneer, and it needed a lot of development. But several options are available that can produce actually good marking strategies. Such as Choke, Stabilized AVQ and other non-linear detection/marking techniques, such as the actual AQM I worked on (Diffusion Early Marking).

  2. Bruno George Moraes Says:

    hi,

    Just wanna point out this other tcp congestion with amazing tests:

    http://media.cs.tsinghua.edu.cn/~multimedia/tcp-fit/

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s


Follow

Get every new post delivered to your Inbox.

Join 609 other followers

%d bloggers like this: