Data point: Annex M ADSL2 can be approximated as 10M down, 2M up in practice. Throw BitTorrent at that, and round-robin delay absolutely is relevant. ADSL1 connections will be even more so. Not everyone lives in a city in Scandinavia.
So a simple tiered scheme which can distinguish VoIP from BitTorrent and both from general traffic, and applies fq_codel to each tier, is a good idea.
- Jonathan Morton
Hi Juliusz,
On Dec 8, 2013, at 14:25 , Juliusz Chroboczek <jch@pps.univ-paris-diderot.fr> wrote:
>>> The promise of fq_codel is that we can get rid of our prioritising
>>> hacks -- if we need that kind of features, then fq_codel has
>>> failed.
>
>> Is that really true? given enough concurrent flows, critical flows
>> might be delayed purely be the round robin scheduling of equally
>> "worthy" packets in fq_codel
>
> At 100 Mbit, one full-size Ethernet frame is 120us. This means that
> if you want your VoIP traffic to have less than 30ms delay, you should
> in principle reach your deadline as long as you have fewer than 250
> congestion-limited flows at a given time.
Well, is 250 enough and are the 250 really realistic in a residential setting? Currently not doing much of anything my router has 142 active connections (according to conntrack) so 250 might be on the low size for a device that routes multiple devices. Plus, unfortunately, most residential internet connections are asymmetric, so on the upload will allow fewer congestion-limited concurrent flows before the round robin delay will impede the VOIP session. (In Germany residential VDSL with 100Mbit/s downlink will run at 40Mbit/s uplink, so hopefully not a big issue, but most cable providers keep the upload below 10Mbit/s, typically 5Mbit/s for 100Mbit/s download). So we talk about an order of magnitude fewer flows required to make phone calls "interesting".
So I still think that for VoIP prioritizing might still be required until supplied minimum bandwidth gets higher.
>
>> so some residual priory system might still make sense...
>
> For throughput-sharing reasons, perhaps. For latency reasons, hopefully not.
Even at 1000 symmetric I still think it would be a good idea to isolate really latency critical traffic from the rest, even if under normal circumstances there should be no problem, I guess a "better safe than sorry" approach. But, hey I do not do this for a living so I might be on the wrong track here.
best
Sebastian
>
> -- Juliusz
_______________________________________________
Bloat mailing list
Bloat@lists.bufferbloat.net
https://lists.bufferbloat.net/listinfo/bloat