[Codel] [RFC PATCH] Codel: Enable packet drop with ECN-marked packets on a threshold

[Dave Taht]

On Sun, Jun 17, 2012 at 11:10 PM, Eric Dumazet <eric.dumazet at gmail.com> wrote:
> On Sun, 2012-06-17 at 22:40 -0400, Dave Taht wrote:
>
>> (side note, I noticed fq_codel defaulted to 10k packets which is
>> rather excessive for tiny routers - I just trimmed that down
>> significantly for cerowrt and the upcoming 3.3.8-4 release has the rfc
>> patch in it)
>>
>> And apologies for not seeing this long ago,
>
> 10k packets is too small to absorb a burst of 64bytes packets on 10Gb
> links. Whole CoDel point is to accept packets at enqueue and drop them
> at dequeue _if_ sejourn time too big. Number of packets should be
> irrelevant.

I agree the setting is good for 10GigE. 40GigE is coming up. 100GigE
is being specified...

> If you don't know how much packet can be sent on wire per unit of time,
> just set a reasonable big limit.
>
> 1000 packets limit is not reasonable, while 10k is.
>
> linux average machines have more ram than tiny routers, dont assume we
> release specialized code. It should be generic enough, granted it can be
> easily tuned.

I don't have a problem with the limit being set to sane values for
modern systems. I note that for 10GigE we are also setting target,
interval at
lower values manually already, so setting that value additionally
seems less work than requiring that systems running at GigE and below
set it - systems that already have a pretty standard 1000 packet limit
baked into their PFIFO_FAST assumptions.

An overall better no knobs approach would be to be able to set the
limit based on the observed hard or soft line rate of the interface.

That said, so long as the need for setting a lower limit in fq_codel
is well documented for smaller systems like tiny routers I don't have
a problem with it. A sane limit can be calculated for soft-limited
interfaces on devices with memory constraints as well

>



-- 
Dave Täht
SKYPE: davetaht
http://ronsravings.blogspot.com/