LARTC - Aug 2007 - Round Robin trafic shapping

I have this problem:

I have an Internet line input with variable speed. I have a max speed and a 
min speed: Vmax and Vmin.

The speed is always changing between Vmax and Vmin. I want to share the 
actual bandwidth (you don''t not how much, you only know the speed is
between
Vmax and Vmin) for N clients. The bandwidth should be shared so nobody can 
get more bandwidth than the others.

There is some queue disciplines like esfq and WRR. But theses ones only work 
if you know the actual bandwidth. HTB works only width fixed bandwidth too.

I have studied all shaping and queue disciplines in linux, and I don''t
have
a real solution yet.


Do you have any idea?

Santiago


-- 
Este mensaje ha sido analizado por MailScanner
en busca de virus y otros contenidos peligrosos,
y se considera que está limpio.
MailScanner agradece a transtec Computers por su apoyo.

I have this problem:

I have an Internet line input with variable speed. I have a max speed and a 
min speed: Vmax and Vmin.

The speed is always changing between Vmax and Vmin. I want to share the 
actual bandwidth (you don''t not how much, you only know the speed is
between
Vmax and Vmin) for N clients. The bandwidth should be shared so nobody can 
get more bandwidth than the others.

There is some queue disciplines like esfq and WRR (w. But theses ones only 
work 
if you know the actual bandwidth. HTB works only width fixed bandwidth too.

I have studied all shaping and queue disciplines in linux, and I don''t
have
a real solution yet.

Do you have any idea?

Santiago
------- End of Forwarded Message -------


--
Open WebMail Project (http://openwebmail.org)


-- 
Este mensaje ha sido analizado por MailScanner
en busca de virus y otros contenidos peligrosos,
y se considera que está limpio.
MailScanner agradece a transtec Computers por su apoyo.

>
> There is some queue disciplines like esfq and WRR (w. But theses ones only
> work if you know the actual bandwidth.

As far as I undestand, at least sfq should work without knowing the actual
bandwith (you don''t need to specify it in the qdisc creation). The
problem
with it is that it only works in the bottleneck of the network, where the
queues form, which is usually the machine which has the hardware that makes
the connection to the Internet. For example in this configuration...

Internet <--- cable/dsl/etc. ---> Router/Modem <---ethernet--->
Firewall
<---ethernet---->LAN

shaping outgoing traffic (traffic to the internet) with only sfq, esfq,
prio, etc... will be useless if it is done in the firewall, becouse the
queues will form in the router. It could be done in the same router if it
has the capability (for example, it runs Linux inside and it is accessible
by ssh or telnet). I haven''t tried WRR but the same should be applied.

Theoretically, at least as I see it, HTB could also be used without knowing
the actual speed in the bottleneck, you just should set the root class to a
speed higher than Vmax and adjust the rates of the leaf classes to the
desired ratios, or playing with the quantums so that it behaves like a WRR
or DRR (which is that it uses internally when the classes are borrowing).
But I have made many attempts in this sense and all of them have been
unsuccessful, don''t know why.

For incoming traffic it is another story, it can be done in the firewall,
but it is a little trickier.

Even more theoretically, outgoing shaping maybe could be done also in the
firewall if the router supports ECN, so that the queues could be formed in
the firewall using this information without nedding to reduce the bandwith,
but this is just a mental experiment which still have pending to brought to
practice, so don''t pay much attention to this...


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

>
>  I am interested in the incoming traffic (from Internet to lan users).
> This is the problem.
>
Then there is no possibility to change the shaping to the bottleneck, as
long as it is in the ISP routers. So you have to artificially create it in
order to have a queue you can shape. There''s only two ways you can
achive
this, either you can:

a) Reduce the speed, the one you already know but are not happy with.

b) Increase the delay, you could achive the queue just introducing a fixed
delay to every packet that goes to the LAN. There is a qdisc called
''netem''
that can do this, in principle it is only intended for emulating and
testing, but it also can be used for this purpose (online gamers will not be
very happy with this solution). I don''t know if there is any other
queuing
discipline that can achive this or if it is not good to do this for some
technical reason.

Maybe it could be possible not to penalize neither speed or delay using some
kind of advanced solution that monitors how much bandwith is consuming each
host and penalize only the hosts that are taking some % more than the
average, or % more than totalacutalspeed/N. But it will be very compilcated
as long as when a host is below the average you don''t know whether it
is for
congestion or just becouse it is not making intensive use of the net.

Good luck, and remember that for incomming traffic you can only shape TCP
connections, it has no sense dropping UDP packets as long as they have no
speed regulating mechanism, and once you have recive them the bandwith has
already been wasted.


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