[Asterisk-Users] Using Linux traffic shaping to prioritise SIP/IAX traffic?

[Alberto Bertogli]

On Tue, Jun 10, 2003 at 09:58:14PM +0200, Emanuele Pucciarelli wrote:
> Il mar, 2003-06-10 alle 20:07, Stephen Davies ha scritto:
> 
> > > When the "tos" option is set correctly (to "nodelay"), the default
> > > queueing in recent kernels already does that, because the pfifo_fast
> > > queue is used (if I recall correctly).
> > 
> > But there is never any queue on my Linux box.  It all storms out of
> > the ethernet interface and gets queued up in my cable modem which
> > doesn't know anything about tos settings.
> 
> That is not entirely correct.  There is an output queue, and pfifo_fast
> is the default (see the LARTC Howto, 9.2.1.1).  But you are right when
> you say you need something to slow down the data;the simplest  choice
> should be addingthe Token Bucket Filter (9.2.2.2).  
> 
> But if the wondershaper already does it all, then it's probably better
> to go with it... :)

You should read further into lartc.org, the linux traffic shaping
capabilities are really wide and you can find lots of ways of doing what
you want.

In your case, I guess the logical choice would be to use HTB, with two
classes, let's say if your cablemodem is 512kbps, you can save 112kbps for
voice and signalling (yes it's extreme but it's an example), and 400kbps
for the data. Also, you can put the former with top priority to minimize
latency.

Under those you can use sfq to make everything more fair, that helps a lot
when saturating.

This is a very short script, I'm sure there is something like this on
lartc or htb's website.

Something like this (completely untested, from memory so don't trust me):

# delete the existing qdisc
tc qdisc del dev eth1 root 2>&1 > /dev/null

# init the htb qdisc
tc qdisc add dev eth1 root handle 1: htb

tc class add dev eth1 parent 1: classid 1:5 htb rate 512kbit
tc class add dev eth1 parent 1:5 classid 1:10 htb rate 112kbit prio 0
tc class add dev eth1 parent 1:5 classid 1:20 htb rate 400kbit prio 2

# sfq for all of them
tc qdisc add dev eth1 parent 1:5  handle 500: sfq perturb 10
tc qdisc add dev eth1 parent 1:10 handle 100: sfq perturb 10
tc qdisc add dev eth1 parent 1:20 handle 200: sfq perturb 10


And now you need to set the filters up, which can be based on tc or using
iptables' MARK target.

For instance:

# everything marked 10 in iptables go to 1:10 (the 112kbit)
tc filter add dev eth1 protocol ip parent 1:0 prio 1 handle 10 fw \
	flowid 1:10

# everything marked 20 in iptables go to 1:20 (the 400kbit)
tc filter add dev eth1 protocol ip parent 1:0 prio 1 handle 20 fw \
	flowid 1:20


And then mark in iptables (i don't remember sip's port very well, and you
should also add RTP stuff too):

# sip, mark 10
iptables -t mangle -A PREROUTING -i eth2 -p tcp --destination-port 5060 \
	-j MARK --set-mark 10
iptables -t mangle -A PREROUTING -i eth2 -p udp --destination-port 5060 \
	-j MARK --set-mark 10

# default, mark 20
iptables -t mangle -A PREROUTING -i eth2 -d 0.0.0.0/0 \
	-j MARK --set-mark 20

In this case you could have used tc's native filters which are much faster
than iptables, but also harder to setup, so if this is your first approach
to this stuff I wouldn't recommend them (and don't worry, the speed
difference is _not_ noticeable for that bandwidth).


I hope this helps, however this is all untested, (ie. just wrote it) so
please look into the docs to find out more.


Thanks,
		Alberto