Re: ns limitations?

[John Heidemann <johnh@ISI.EDU>]

On Wed, 21 Apr 1999 18:58:24 MDT, Kristin Wright wrote: 
>
>Hello,
>
>Can anyone provide concrete examples of a computationsl 
>limitation in ns or approximations of limitations based on 
>empirical evidence? For example, has anyone had to limit 
>the number of nodes in an experiment based on processing time? 
>
>We're interested in including ns in a network testbed proposal. 
>However, an "inherent limitations of simulators" argument 
>continues to surface. After some discussion, it's clear that
>none of us have a *concrete* idea of where the limitations
>might be. We can imagine where some might exist, but we
>cannot attach numbers to the scenarios. 
>My own experience with ns has not revealed any
>such limitations. 
>
>(I've already pointed my colleagues to the ns 
>publication web page for evidence of what simulations 
>*can* be run.)
>
>Thanks very much for any help. 

There is much discussion of this in the ns-users archives (look for
posts by Polly Huang).  Typically we've seen simulations limited more
by memory than by processing time.  We've attacked this bottleneck
with low-level optimizations (often very effective) and the use of
abstraction techniques to avoid unnecessary detail in simulations when
it's not needed.  See Huang's MASCOTS paper (available on the
``research using ns'' web page) for descriptions of some of the
abstraction techniques currently available.  She's also done large
detailed simulations---Polly, can you fill in the details about
numbers of clients/servers/connections?

I don't quite understand what you mean by the ``inherent limitations
of simulators'' argument.  If you mean that they only implement the
details considered important by their implementers, that's certainly
true.  I think you'll see an increasing level of work to answer the
question of if they have enough detail for a given experiment (with
answers going both ways for different experiments).

But if you mean some limit in terms of scalability, I don't
understand.  Physical testbeds have the same problem (limited numbers
of nodes and links), and I'd argue it's worse because each node or
link costs real money (sometimes lots of it for real cross-country
links) rather than being completely virtual.

   -John Heidemann