[e2e] [Fwd: RED-->ECN]

[Steven Low]

>         AQM in the absence of ECN can be counter productive. If you try to
> control the buffer to low levels, then you reduce delays. However that
> means that loss must go up almost quadratically (taking the
> simplified throughput equation for TCP, sqrt(p)*RTT = Constant since link
> capacity is not changing). That increases the probability of redundant
> retransmissions as also the probability of flows going into
> timeouts. Goodput suffers. Hence, decoupling congestion measure and
> performance measure may not help if your feedback mechanism is drops
> instead of marks, and will in all likelihood hurt performance. Thus, ECN
> is *critical* for the success of AQM schemes.
> 
> -Vishal

Hi Vishal,

I completely agree that ECN would be great for AQM.  I agree also with
your observation that larger delay in fact reduces loss rate.   However
I'm not sure if I agree that AQM without ECN *necessarily* leads to
worse
performance.  I think it probably depends on 1) various parameters of
the
scenario, and 2) the specific AQM scheme.  Our preliminary ns-2
simulations
with REM actually shows improvement in goodput and delay over DropTail
(using NewReno) both with or without ECN (see preprint "REM: Active
Queue
Management" on our website at the end of this email).  I can imagine one
can probably design scenarios where what you describe is the case.  An
interesting question is to characterize the conditions under which AQM
without ECN helps or hurts.  

Another question is whether we should rely on large (queueing) delay to 
keep loss rate down?   This doesn't seem a good idea, nor necessary.

Steven
ps. We hope to post our ns-2 scripts so that anyone interested can
play with REM.
__________________________________________________________________
Steven Low, Assoc Prof of CS & EE 
slow at caltech.edu			netlab.caltech.edu
Tel: (626) 395-6767			Caltech MC256-80
Fax: (626) 792-4257			Pasadena CA 91125