A few questions

[Muneyb Minhazuddin <Muneyb.Minhazuddin@tip.csiro.au>]

Hi Folks,

Reporting from the Olympic city, your programmable network correspondent.

I am going to try an experiment on my Accelar Programmable network
testbed. I will first attempt to explain what I am going to try and
propose a procedure to do it, I would like your feedback as a sanity check
in case I am assuming the wrong things about the accelar's capability.

My network looks like this, it already exists, however all the routers are
linux based, and I have successfully experimented this scenario on my
pure linux testbed. 

		Smartbits
		   |
                   |
		   V
 ----------    -----------    ----------   -----------   ----------
 |  VOD   |    |         |    |        |   |         |   | VOD    |
 | Server |--->| Ingress |--->| Core   |-->| Egress  |-->| Client |
 |        |    | Router  |    | Router |   | Router  |   |        |
 ----------    -----------    ----------   -----------   ----------
		   ^
		   |	
		   |
		Voice 


The Video On Demand is an MPEG-2 adaptive hierarchically coded application
written in-house, it has different versions UDP based, TCP based, over
IP or over ATM, incorporated with RTP and RTCP.

For this experiment I am using the UDP based version over IP. Basically
there are 3 streams B frames, P frames and I frames running on three
different UDP ports. Under conditions of congestion created by say voice
traffic which is of very high priority, the B frames that are of least
importance are to be dropped automatically by the Ingress router 
(this being the part of SLA that the customer has agreed to). As trivial
as that.

My Linux Implementation
-----------------------

1/ In the core router, I have a complete diffserv enabled mechanism i.e a
Expedited Forwarding + Assured Forwarding + Best Effort capable router. It
is implemented using Class Based Queuing with set of traffic control
filters and different queue mechanisms.

The queue structures are 

*EF packets going through a short FIFO queue
*AF packets going through four GRED queues ( accounting for the four AF
classes). Each GRED queue in turn is equipped with 3 RED queues to provide
for the 3 Drop Precedences.
*Best effort packets go through a RED queue.

The voice traffic is targeted for EF.
The VOD traffic is ideal for AF.
I have background traffic generators (Smartbits) to congest the network.

2/ I have traffic control filters on the Ingress router. They have
forwarding rules set for the 3 UDP ports. I also have a program, wrapped
around libpcap which captures and analyses all the RTCP packets flowing
back from the client to the server. This program monitors the Frame Loss
Percentage being reported by RTCP for B, P and I frames. When the frame
loss exceeds unacceptable levels (which arise due to congestion) then the
filters on that stream are changed automatically to drop all the packets
of that particular stream.

With the above setup, I start the VOD application, then some background
traffic. There are no losses, all my traffic gets through. Now I slowly
ramp up my voice like traffic which has a significant allotment of
bandwidth to have an effect on the VOD and completely stop any background
traffic. As losses in the B Frames are encountered (as they are very big
in size, Refer MPEG RFC), my program writes in the new filters which drops
all the packets of the B frame stream. 

Therefore the network has reacted to congestion based on the SLA with the
customer.

Using the Accelars
------------------

1/ Ideally

In my network I would like to replace all my linux routers with
accelars. 

Procedure
I intend to write an oplet that uses JFWD in the ingress router
to write the traffic filters. It also uses JCapture to capture the RTCP
packets. When the frame loss percentage exceeds unacceptable levels then
JFWD will write in the appropriate traffic-filters whose Action will be to
drop the packets. 

Q1. Is this the right way to proceed?

Q2. (To Rob Duncan) Are JFWD and JCapture ready for use? You mentioned in
a previous mail that you are testing a new release of JFWD?

2/ Practically 

We have only 2 Accelar 1100B's. From my experience, it has only 2 queues
and priority can be enabled by Port, VLAN, Mac address, IP Flow, Ip
filters, etc. So I might have to just consider Voice and VOD applications.  

As I can't have so many queues I will have to use the accelar only at the
Ingress for now. 

Q3. Can my procedure, as mention above, possible on an Accelar at the
Ingress?


I understand that the 8600 has 8 queues (from an old email conversation
with Doug Moeller). 

Q4. Will I be able to replace my Core router with the 8600 ?


Q4a. How do configure my Diffserv capable router with the following queues

(repeated here from above for convenience)

CBQ based schedulers with
 
*EF packets going through a short FIFO queue
*AF packets going through four GRED queues ( accounting for the four AF  
classes). Each GRED queue in turn is equipped with 3 RED queues to provide
for the 3 Drop Precedences.
*Best effort packets go through a RED queue.

Q4b. Or their equivalent mechanisms? 

I understand that you have WRR mechanisms for scheduling, but are their
RED like queues (not relative priority queues). 

I am basically looking for parameters like these, which I use in my Linux
implementation.

limit - actual physical size of the queue.
min   - Minimum threshold in Kilobytes.
max   - Maximum threshold in Kilobytes.

where,
min < max < limit
 
(probability)
|      /(1)   /(2)               /(3)
|     /     /              /
|    /    /          /
|   /   /       /
|  /  /     /
| / /   /
--------|------------------------|------------>
        (min for eg.3)           (max for eg.3) 
 
I will try to give examples from the above crude graph :-)
 
1. limit 20KB min 5KB max 10KB, slope indicates small delay through the
queue however drops will happen earlier than the other cases
 
2. limit 30KB min 8KB max 24KB, 
 
3. limit 100KB min 15KB max 75KB
 
Delay through the queue goes up cases 1-->3. However drops will happen
earlier in case 1 than in case 2, in case 2 before case 3.

>From my understanding of the AF RFC, RED is quite key for its
implementation (please refer Diffserv mailing list and Diffserv
Implementation mailing list that I co-host with Brian Carpenter).

Any answers to my questions would be appreciated.

Cheers
Muneyb.
_________________________________________________________
Muneyb Minhazuddin - Telecommunications Research Engineer    
CSIRO Telecommunication and Industrial Physics                
Sydney, NSW, Australia.                                    
                                                             
Phone no. : 61 2 9372 4113
FAX       : 61 2 9372 4490          
e-mail    : mminhazu@tip.csiro.au  
Home Page : http://www-networks.tip.csiro.au/~mminhazu
---------------------------------------------------------