Re: [aqm] Status of the GSP AQM?

[Jonathan Morton <[e-mail]>]

On 15 Dec 2017 19:19, "Francini, Andrea (Nokia - US/Murray Hill)" <
[e-mail]> wrote:
>
> If “conjunction with flow isolation” means combination of the algorithm
with a flow queueing arrangement, there is logically no restriction in
realizing it. We tested FQ-GSP on ns2, getting similar results as with
other FQ-AQM schemes (never worse, never overwhelmingly better in the
scenarios we looked at). Since the algorithmic simplicity is not as
critical in lower-speed links, we thought there was little value in trying
to add one more scheme to an already crowded space.

Perhaps a better wording would be, "it seems very unlikely to outperform
existing AQMs in conjunction with FQ".

> Also (and this is just my opinion), I don’t think that combining FQ and
AQM is a good idea, because it imposes a single policy on all flows despite
the variety of their needs. I like a plain FQ with large buffer much
better, because it guarantees bandwidth fairness and makes every
application solely responsible for the queuing delay it gets.

I have a couple of compelling arguments in favour of an AQM-FQ combination:

1: ECN aware flows are able to avoid incurring packet losses, by receiving
and acting on information about link congestion when they reach their
fair-share throughput.  That should be music to the ears of those industry
members who are apparently allergic to packet loss.

2: When it is unfortunately necessary to put the shaper downstream of the
physical bottleneck link (ie. when consumers must compensate for an ISP's
bufferbloat), AQM is the only reliable way to keep the bottleneck queue
empty.  Combining it with FQ ensures that only the saturating flows receive
congestion signals and potentially lose packets.

- Jonathan Morton