[e2e] Why do we need TCP flow control (rwnd)?
Lachlan Andrew
lachlan.andrew at gmail.com
Fri Jul 11 10:08:14 PDT 2008
2008/7/11 Christian Huitema <huitema at windows.microsoft.com>:
>
> The point about nothing else is certainly valid. If you have to perform your computation using log tables and an actual spreadsheet, then tackling anything besides Poisson will be very challenging. However, it has been known for some time that Poisson was at best a first degree approximation, even for phone traffic.
I disagree. A time-invariant Poisson model (a la M/G/1) may be
very inaccurate, but useful statistical analysis can be done with time
varying Poisson models, which capture diurnal variation and flash
crowds. (No, that is not the same as saying "any two points can be
fitted by an exponential", as was previously said.)
As Mark Crovella correctly pointed out, *session* arrivals (analogous
to phone calls) seem to be well modelled by a Poisson process, even
though individual connections are not, and packets certainly aren't.
We shouldn't throw out all models which use the word "Poisson",
because *some* Poisson models are inaccurate. Isn't that like
throwing out all measurement studies on the grounds of being
"anecdotal evidence"?
> Even the duration of phone calls was probably not well modeled by time independent departure processes
Of course. It has long been known that call holding times are far
from exponential. However, for an M/G/K/K queue (the one used for
telephony) behaves exactly the same as M/M/K/K by several important
measures such as blocking probability, and so there was little to be
gained by studying the more complicated model.
Cheers,
Lachlan
--
Lachlan Andrew Dept of Computer Science, Caltech
1200 E California Blvd, Mail Code 256-80, Pasadena CA 91125, USA
Ph: +1 (626) 395-8820 Fax: +1 (626) 568-3603
http://netlab.caltech.edu/lachlan
More information about the end2end-interest
mailing list