[Cerowrt-devel] 10gige and 2.5gige

[Toke Høiland-Jørgensen]

Sebastian Moeller <moeller0 at gmx.de> writes:

> Mmmh, I guess our approach at traffic shaping does not scale well at those speeds. Maybe this could be fixed with larger batching?
>
> I think it might be worth trying to switch to simple.qos/fq_codel and
> set a somewhat larger burst/quantum duration in defaults.sh, then
> disable BQL on the NIC and configure a beefy txqueuelen on the
> interface. This might help making SQM limp along to higher rates. If
> that would actually work, we could try to see whether we can make cake
> learn coarser batching ("bursts" and/or quantum) at high rates (but I
> did not check what cake does internally, it might already do this,
> @jonathan?)....

I fear the problem is rather the lack of multithreading. I have a fairly
beefy 8-core ARM box for my main router these days, and even that can't
forward at a gigabit on a single core. There was a bug in the HW
configuration so all traffic was sent to a single core, which resulted
in 50+ms of bloat and traffic capping out way short of a gigabit. Now
that it's fixed and traffic is mixed over all eight cores I have smooth
sailing and no bloat. Thankfully I don't need to shape, so I'm just
running straight mq+fq_codel on the physical interface....

-Toke