Re: [LARTC] pfifo_fast as leaf qdisc for htb classes

[Andy Furniss <andy.furniss@dsl.pipex.com>]

panca sorin wrote:
> Thank you Andy! I did't find why I could't use the fw
> match in ingress. Now it works.

It will work on marks set in prerouting if you don't select packet 
action in your kernel config (near policer).

> 
> Now I ran into another problem:
> When I try to add a pfifo_fast leaf qdisc dor htb
> classes it seg.-faults. Is pfifo_fast a special qdisc?
> I can succesfully add other type of qdiscs.
> 
> I found that a similar problem was reported in
> february on a MIPS architecture.

Yes I think it's special - but the default behavior of prio is much the 
same.

> 
> root@zefir:/usr/src# tc qdisc add dev eth0 root handle
> 1: htb default 10
> root@zefir:/usr/src# tc class add dev eth0 parent 1:
> classid 1:1 htb rate 100Mbit ceil 100Mbit burst 50k
> cburst 150k

FWIW if you are shaping 100meg ethernet those settings would be too high 
I would use -

... rate 99mbit ceil 99mbit mpu 84 overhead 38 ...


> I think htb and pfifo_fast assures a better QOS than
> htb & sfq because I can priorize interactivity.
> 
> Thank you in advance for any sugestions.

Yes use prio if you are going to give one class for each user, you will 
still have to set the tos bits/classify yourself as you can't rely on 
apps/inbound having them set properly.

Also remember that if you are shaping inbound traffic you need to back 
off from the link rates more than outbound so that a queue can build up. 
If the link is slow then you'll need to keep the queues short aswell, so 
you drop packets.

Andy.

_______________________________________________
LARTC mailing list
LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc