[e2e] Re: queue averaging introduces delay

[Dennis Ferguson]

Hello,

>  From queuing theory, only when the sources have a Poisson arrival rate 
> does the length of the queue directly relate to the number of sources 
> (and thus the appropriate congestion response).
> 
> Both the TCP request arrival rate and the inter-packet arrival rate of a 
> single TCP connection are decidedly non-Poisson.  This makes placing 
> queue length (and especially instantaneous queue length) at the heart of 
> AQM schemes dubious.  
> 
> Decoupling queue length from AQM schemes can provide significant 
> improvements in packet loss, queuing delay, and link utilization.  If 
> you try setting w_q in RED to something extremely small or try using 
> Blue QM, you can see this difference directly.....

I may be missing something obvious, but it is hard for me to see how
an AQM scheme which makes adjustments to its control parameter when
the queue length is instantaneously zero (a "link idle event") or
when the queue length is instantaneously beyond some limit (a "packet
loss event") could be said to be decoupled from the instantaneous
queue length.

I am, however, sympathetic to the implied concern about what it is
which AQM schemes pick as control parameters.  I'd personally prefer
not to have to pick a target average queue length, since this seems
of only secondary relevance.  I think I'd instead prefer the
control algorithm to keep the output circuit as close to 100% full
as possible under sustained congestion, with the control algorithm
on its own finding the minimum average queue length for which full
output circuit utilization can be maintained.  This, of course, may
require magic.

I'd also note that most AQM proposals seem to leave half of the current
problem on the table.  Most routers now implement multiple output queues
per circuit in support of differentiated services.  This means that, from
the point of view of any individual queue, the queue drain rate is not
necessarily a constant but rather depends on the amount of traffic through
the other queues.  Since most analyses seem to assume that there is a
simple fixed relationship between queue length and queuing delay, it
leaves a non-trivial exercise to the reader to determine the applicability
of the results to situations where this relationship is violated.

Dennis Ferguson