Re: [aqm] References on AQM test results and fq_nocodel

[Toke Høiland-Jørgensen <[e-mail]>]

David Lang <[e-mail]> writes:

> if the fq portion is being gamed, how severe can the imbalance be? Is
> it a matter that if there are N flows without gaming the system, and
> each is getting 1/N bandwith, then if a cheater uses M flows the
> cheater gets M/(N+M) of the bandwidth?

Yes, not counting hash collisions, that would be the case.

> more importantly, is there any way short of a massive number of flows
> (essentially a DDOS), that this gaming of the system can hurt the
> latency of the other flows? or is it just the relative bandwidth of
> the different apps that suffers, and if an app is low bandwidth, but
> latency sensitive (i.e. VoIP), it won't be affected until it's "share"
> of bandwidth drops below the minimum it needs?

Well, since fq_codel is hash based, if you manage to hit the same bucket
as the VoIP flow, that VoIP flow is not going to get the nice implicit
priority it would otherwise from the flow queueing. So yeah, if you,
say, send out 1000 packets at once on different UDP ports, then the VoIP
flow is going to notice. But in this case you're basically DOS'ing the
router; and FIFO queueing and pure AQMs fare pretty badly as well...

> Looping back to the start of the topic, the slides apparently are
> being read that fq without codel is just as good, or better than
> fq_codel, so codel can be dropped, and only fq is needed.
>
> Is this what you are concluding from this? Or is that a misreading of
> something (the data, your work, or me misreading this thread)?

Well, *for the tested scenarios and metrics* that is true. However,
there can be other scenarios where you will benefit from having AQM on
each of the queues. Several such hypothetical scenarios have been
mentioned in this thread. The open question, as far as I'm concerned, is
*how much* a difference it makes. I.e. exactly how much the performance
degrades if it is absent, and in which scenarios. As I said before, I
don't think anyone has ever actually tried it and produced data...

-Toke