- Stefan
> -----Original Message-----
> From: dovrolis@hertz.ece.wisc.edu [SMTP:dovrolis@hertz.ece.wisc.edu]
> Sent: Thursday, May 14, 1998 8:54 AM
> To: jacques.kerberenes@cnet.francetelecom.fr
> Cc: end2end-interest@ISI.EDU; ippm@advanced.org
> Subject: RE: Current use of RED?
>
> Jacques,
>
> > But simulation tools can only simulate a few number of TCP
> flows.
> > Are the results still valid with realistic
> > traffic patterns : thousands or dozen of thousands flows ? We
> don't
> > know. I remember a contribution
> > of Pr Raj Jain at the Vancouver ATM Forum meeting about TCP over
> > ABR. We asked him if his results were still valid for thousands of
> > concurrent TCP flows and he answered : I can't say it.
>
> Getting slightly out of the main theme of this thread, I can't
> agree more with you, or with Prof.Jain. There are some very
> interesting things that occur in core routers that simulations
> (at least the current simulations) cannot show.
>
> A "classical" example: Suppose that you do simulations with a
> dozen of TCP connections to find out the appropriate buffer size
> in the output interfaces of your router. It is quite possible
> that you will end-up with the well-known result: "Make it
> equal to the bandwidth-delay product, for a `typical' round-trip delay
> and for a `typical' bottleneck-link bandwidth". The point is,
> however, that in a core router serving 10000's of simultaneous flows
> this buffer size will be probably too small. You should instead
> make the buffer size proportional to the maximum number of flows that
> can be serviced at any time, as in the worst-case, all of these
> flows would like to have at least one packet in the queue of that
> interface. I know that I oversimplify, but the point that I try
> to make is: Simulations with a dozen of flows will not show you
> this effect.
>
> It is scary to think that all the nice ideas that people
> suggest in the literature are based on this kind of "small-scale
> simulations".
>
> IMHO, it is time for the network research community
> to start demanding "large-scale simulations" in any proposed
> new scheme, before considering it correct or viable. In parallel,
> a very interesting research area that people may want to look
> at is: traffic-models and simulation mechanisms for
> large aggregates of TCP traffic. This is rather mandatory,
> as not many network researchers have access to supercomputers,
> or want to wait for weeks for a simulation to finish.
>
> Constantinos
>
> PS: I am not sure that this thread should go to ippm as well.
> I think so. Am I right Vern?