[LARTC] Re: tc questions

[Andy Furniss]

Alejandro Ramos Encinosa wrote:
> On Tuesday 10 April 2007 22:11, Andy Furniss wrote:
>> Alejandro Ramos Encinosa wrote:
>>> tc qdisc add dev eth1 parent 1:20 handle 120: sfq perturb 10
>>>
>>> tc class add dev eth1 parent 1:20 classid 1:21 htb rate 49mbit
>> This is a misconfiguration, it doesn't make sense to add sfq and another
>> htb class to 1:20.
> ...why? The case I am trying to deal with is an scenario where some traffic 
> goes into 1:20 (something like the traffic from/to the subnet 10.6.70.0/24) 
> and then, I want to shape specifically some other traffic type (for example, 
> the ssh connections from/to subnet 10.6.70.0/24). Is there another way to do 
> it? Please, take a in mind that (in my example) I want to enclose the whole 
> traffic from/to the subnet 10.6.70.0/24 and from that traffic I want to give 
> an special treatment to ssh traffic.
>> Andy.
> Regards, Ale.
> 

You could add two htb classes under 1:20 and give one higher prio, or 
you could use the prio qdisc. If you really care about latency and have 
many bulk classes on a slow link then hfsc is better than htb. Linux 
hfsc could still be improved.

sfq and b/pfifo should be added on leafs, so you could still use them if 
you created two classes under 1:20.

If you don't specify a qdisc on htb leafs you get pfifo - but the queue 
length will be chosen from the interface that htb is added to - 1000 for 
eth (possibly too long) or 3 on ppp/vlan (too short), so it's worth 
thinking about queue lengths, adding a qdisc and using the limit 
parameter common to b/pfifo and sfq. (default sfq is 128).

Andy.