Re: ns performance

[Christoph Haenle <chris@cs.vu.nl>]

> Hi,
> 
> I want to use ns to simulate a network with about 5000 multicast nodes
> and about 10000 multicast messages. Can someone please tell me if such a
> large network can be simulated with ns and how long one simulation of
> this scenario approximatly runs (order of seconds, hours or days). I
> want to use a Pentium-150 Computer with Linux or Windows NT.
> 
> Thanks for your help,
> Christian

Hi Christian,

5000 nodes (with approx. 5000 links) alone don't pose the real problem, but 
you might want some losses to occur in your network, I guess. If so, then you 
could configure each link to randomly drop packets, or you could simulate 
packet losses by buffer overflows in the routers. In the latter case, you'll 
need thousands of TCP connections, which will take a lot of memory, which will 
make your scenario utopic on a pentium-150 (unless it comes with GB's of main 
memory).

I've run several tests on an Ultra Sparc II, 2x296MHz, 1280 MB main memory, 
the largest consuming approx. 1.7GB (main and virtual) memory. I could see two 
bottlenecks: The calculations of the routing tables (if precalculated) took in 
my 1300 node scenario ca. 1 hour. I'll become much worse in a 5000 node 
scenario. Further, a lot of state information (caused by prune messages) is 
kept if you have multiple multicast senders (such as in SRM), cause every 
source spawns a separate (pruned) multicast tree.

My simulations weren't particular restricted in time (the longest being 36 
hours or so), but (main) memory requirements were high and when it came to 
swapping, the simulations were slowed down by a factor of 10...100, which 
rendered them almost impractical to complete.

    -Chris.