Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: [RE] Comparing graphical analysis and analytical


As long as the resulting traffic is compliant with it's bandwidth reservation, i.e. it doesn't send more packets per shaping period than reserved, it shouldn't produce any additional latency.

Having said that: If CBR traffic hicks up and then tries to catch-up with the streaming rate, it would violate its bandwidth allocation and indeed would produce additional latency. Is this what you mean by hick up? I assume in my calculation that this does not happen.


"James, David V" <>

11/03/2005 03:02 PM

Maksim Azarov/SMSC@SMSC, <>
RE: [RE] Comparing graphical analysis and analytical

From a worst-case perspective, CBR traffic with a hickup at the wrong time
(traffic stops momentarily) gives much worse latency numbers.

Have you thought about how this could be included in your model?
Its at first counter-intuitive: transients that reduce the loading actually
can make the worst case latencies _longer_.
If this effect could be included in your model, then I would feel more
confident that we are addressing the true "worst case" bandwidths.

From: Max Azarov [mailto:Max.Azarov@SMSC.COM]
Thursday, November 03, 2005 10:15 AM
[RE] Comparing graphical analysis and analytical

I've looked through the David's charts and figures in the effort to synchronize our results on the worst-case latency figures. Few observations:

1) Graphical analysis is considering both input-queue and output-queue routers. My work exclusively deals with the output-queue routers.

2) In graphical analysis of traffic patterns I couldn't find an explicit value for bandwidth shaping period under consideration. This definitely has an effect on the outcome. For the 3 port output queue analysis it seemed that 125 mks shaping period was used.

3) For the output-port router graphical analysis for 3 ports, figures are actually in a very good agreement with the analytical formula for the shaping period of 125 mks and 75% link utilization. In fact when I graphed latency values from the David's paper they where almost linear and also fit tightly with the graphs produced analytically. On the attached graph, analytical curve is black, while curve from the values in the Dave's paper (, table F.3 on the page 10) is blue.

All this leads me to believe that there no drastic disagreement in results at all!