[aqm] AQM schemes: Queue length vs. delay based

[Preethi Natarajan <[e-mail]>]

Hello AQMers:

Just wanted to bring up the following item for discussion either as part of
the recommendations draft or the evaluation criteria during Friday's
session/mailing list.

In access, edge and core routers the draining rate of a queue is affected by
traffic on other queues and thus can vary a lot (depending on the deployment
and traffic conditions). A queue length based AQM scheme such as RED or
derivatives tries to maintain the average queue size around a predictable
value under these changing draining rates. However, this queue size
translates to high queuing delay under low draining rates and vice-versa.
The unpredictability in resulting queueing delay was one of the reasons why
we opted PIE to be a latency-based scheme.

A queue length based AQM scheme could be perfectly valid for certain
deployments. For deployments where predictable queuing delay is expected
under varying draining rates, a latency based AQM is critical. We believe
this should be brought about in discussions somewhere at AQM ‹ perhaps in
the recommendations draft or w.r.t evaluation criteria.

Thanks,
Preethi (on behalf of PIE team)