LARTC - Jun 2003 - Basic principle of enqueunig/dequeuning packets & control traffic flows

Hi people!

Maybe its messages is offtopic but... I have no good ideas how to solve my
problem.

I am writing program that must realtime control traffic flows ( 
control/account/bandwitdth/etc   ip traffic ).
I am using  2.4.x Netfilter queue kernel  module (ip_queue)  so  program can 
get packets via libipq.
In this case I can easy control traffic (set  verdict  NF_ACCEPT or NF_DROP) and
statistics traffic (via counters).
But how to shape traffic flow?
I  wrote very simple wrapper  library (just based on iproute2 sources ) for  tc
command so pogram may set
qdisc, classes and filters on some device where I what  control outbox traffic.
But this is not  convenient way.

How to build own queue(s) for specific traffic flows that going via my program
and setup on them some shaping
discipline like TBF - just for shape traffic, set certain rate?

What is better for system  perfomance when use such shaping and  when using
kernel QoS?

Any suggestions, ideas  are welcomed.

THANKS IN ADVANCE.

PS sorry for my poor english.


-- 
Andriy T. Yanko
wireless@wireless.org.ua

*   Avoid The Gates of Hell use Linux
**  Wenn ich sterbe, legt mir eine Birne in mein Grab ...

_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/

: Maybe its messages is offtopic but... I have no good ideas how to solve
 : my problem.

It''s not at all off topic for this list.  You may find, if you need
more
libipq help, that the netfilter-devel list is a better list, but for this
general "and traffic control, too" question, you have reached the
right
place.

 : I am writing program that must realtime control traffic flows (
 : control/account/bandwitdth/etc ip traffic ). I am using 2.4.x Netfilter
 : queue kernel module (ip_queue)  so program can get packets via libipq.
 : In this case I can easy control traffic (set verdict NF_ACCEPT or
 : NF_DROP) and statistics traffic (via counters). But how to shape
 : traffic flow?

Well, this could be tricky depending on your expectations.  If you know
exactly what sort of shaping you intend to use, and it doesn''t need to
be
dynamically manipulated, then I have an answer for you.

If, on the other hand, you wish to dynamically alter the classes and
bandwidth allocated to a stream, you have a bit more of a task ahead of
you.

[...snip...]

 : How to build own queue(s) for specific traffic flows that going via my
 : program and setup on them some shaping discipline like TBF - just for
 : shape traffic, set certain rate?

I''m no whiz-kid with the libipq stuff, but if you can mark the packets
in
any way, (-j MARK --set-mark $MARK), then you can easily use a static
traffic control/shaping set of qdiscs.

You mention TBF--this is fine and dandy, but doesn''t allow you to
classify
packets.  You may wish to consider learning HTB.  You could set up a large
number of classes each with a specific ceiling.  Then you get token bucket
filter for free on a subset of your traffic--each subset independently.

 : What is better for system perfomance when use such shaping and when
 : using kernel QoS?

So, in summary, I would recommend creating an HTB tree, and then using
dynamically altered MARKs to select packets for a particular output class.
This way, you''ll not need to alter the traffic control setup once you
have
it configured the way you like it.

 : *   Avoid The Gates of Hell use Linux
 : **  Wenn ich sterbe, legt mir eine Birne in mein Grab ...

Warum ausgerechnet eine Birne?  Und wessen?  :)

-Martin

-- 
Martin A. Brown --- SecurePipe, Inc. --- mabrown@securepipe.com

_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/

On Wed, 18 Jun 2003 10:58:52 -0500 (CDT)
"Martin A. Brown" <mabrown-lartc@securepipe.com> wrote:
>  : I am writing program that must realtime control traffic flows (
>  : control/account/bandwitdth/etc ip traffic ). I am using 2.4.x Netfilter
>  : queue kernel module (ip_queue)  so program can get packets via libipq.
>  : In this case I can easy control traffic (set verdict NF_ACCEPT or
>  : NF_DROP) and statistics traffic (via counters). But how to shape
>  : traffic flow?
> 
> Well, this could be tricky depending on your expectations.  If you know
> exactly what sort of shaping you intend to use, and it doesn''t
need to be
> dynamically manipulated, then I have an answer for you.
OK. What  is the answer in this case?
Just set  HTB qdisc with classes on certain device and setup filter 
based on u32 or fw?
> If, on the other hand, you wish to dynamically alter the classes and
> bandwidth allocated to a stream, you have a bit more of a task ahead of
> you.
yep.
> I''m no whiz-kid with the libipq stuff, but if you can mark the
packets in
> any way, (-j MARK --set-mark $MARK), then you can easily use a static
> traffic control/shaping set of qdiscs.
OK . I wrote very simple library that is  wrapper to tc command.
So My program may setup qdisc, classes and filters.
But how to get current rate (of traffic flow)  from certain qdisc  with
associated with certain qdisc ?
In order to  get current rate value of certain class I must use rtnelink
mechanism?
oh... it''s nightmare for me - tried.
> So, in summary, I would recommend creating an HTB tree, and then using
> dynamically altered MARKs to select packets for a particular output class.
> This way, you''ll not need to alter the traffic control setup once
you have
> it configured the way you like it.
but I what change rate  of class everytime... :(
> Warum ausgerechnet eine Birne?  Und wessen?  :)
Einfach.

-- 
Andriy T. Yanko
wireless@wireless.org.ua

*   Avoid The Gates of Hell use Linux
**  Wenn ich sterbe, legt mir eine Birne in mein Grab ...
_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/