In message <[email protected]>
David Collier-Brown writes:
 
> Yup!
>  
> Loss is a bit of like hitting someone with  a 2 x 4" as a clue-stick. It 
> hurts, but it sure does work!
>  
> --dave

ECN might be a touch gentler.  But loss on the order of 10^-4 or less
doesn't hurt too bad.

With classsic TCP (ie: newreno or older and anything newer but not
hyper-aggresive) the estimates of the amount of loss per D*BW product
done by Matt Mathis still hold.  Now if I could just remember what
that estimate was ... in any case you get to extremely tiny loss with
anything over a few Mb/s and it goes up with square of D*BW.

Hyper-aggresive TCP (ie: CUBIC) need to be hit a lot harder.  This is
where FQ/SFQ really comes in handy.

Curtis


> On 02/27/2015 09:16 PM, KK wrote:
> > This could also be done by not having to depend on "loss" as the primary
> > source of feedback...
>  
> On 2/27/15, 6:00 PM, "David Collier-Brown" <[email protected]> wrote:
>  
> > IMHO, removing latency is the aim of FQ. Once done, buffer sizes can be
> > unbounded (save by price (;-))

_______________________________________________
aqm mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/aqm

Reply via email to