[Iccrg] Fwd: [conex] More precision than 1 bit ECN (was: Re: WG energy)

[michawe]

I'm taking the liberty of forwarding this to ICCRG because it is as relevant to the group as it gets, I suppose...

Begin forwarded message:

> From: Bob Briscoe <bob.briscoe at bt.com>
> Date: March 9, 2011 12:30:25 PM GMT+01:00
> To: Caitlin Bestler <cait at asomi.com>
> Cc: conex at ietf.org
> Subject: [conex] More precision than 1 bit ECN (was: Re: WG energy)
> 
> Caitlin,
> 
> Still catching up on parts of past threads...
> 
> At 04:41 23/02/2011, Caitlin Bestler wrote:
> 
>> On Feb 17, 2011, at 10:04 AM, John Leslie wrote:
>> 
>> > Caitlin Bestler <cait at asomi.com> wrote:
>> The other area where convergence would be useful is coming up with an abstract way to report "congestion
>> detected" with more precision than just an ECN bit, so that L4 can utilize this information wihtout having to
>> learn the specifics of L2 congestion control protocols.
> 
> This as relevant to tsvwg & iccrg as to ConEx, but I'll keep it to ConEx for now...
> 
> Matt Mathis taught me something important in respect of how many bits you need per frame/packet for congestion control. The critical metric is the amount of signalling information per *window*, not per packet/frame. You can maintain the same amount of signalling information per window with just 1 bit per packet as long as the congestion control keeps the packet rate proportional to 1/p, where p is the marking probability.
> 
> Intuition: with a window of 100 frames, if the marking probability is 1% you will get one mark per round trip. If when the marking probability reduces /100 to 0.01% you speed up x100 to 10,000 frames per window, you will still get one mark per RTT. So you will still have the same rate of control information per time.
> 
> See slide 8 of this Anaheim ICCRG presentation jointly from Matt & I:
> <http://www.bobbriscoe.net/present.html#1003iccrg>
> 
> 
> DCTCP, Relentless TCP and Scalable TCP all aim for 1/p (this is why Scalable TCP was called scalable). TCP NewReno's window is proportional to 1/sqrt(p), but more modern algorithms like these are moving closer to 1/p.
> 
> 
> Bob
> 
> 
> ________________________________________________________________
> Bob Briscoe,                                BT Innovate & Design 
> _______________________________________________
> conex mailing list
> conex at ietf.org
> https://www.ietf.org/mailman/listinfo/conex

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://oakham.cs.ucl.ac.uk/pipermail/iccrg/attachments/20110309/96dd473f/attachment.html