[Beowulf] Gigabit Eithernet Ring

[Daniel Pfenniger]

Hi Steve,

Your site was unavailable, so I cannot comment on your N-body method.  But if you use a 
communication ring you might well be calculating the full N(N-1)/2 interactions, which 
is computationally very expensive and not necessary for many N-body problems.  Tree 
based methods lower the computational complexity to O(N lnN), or even
O(N).  The current most efficient O(N) Fast Multipole like algorithms is by Walter 
Dehnen (http://arxiv.org/abs/astro-ph/0003209).  Such calculations are approximate but 
with a tunable accuracy parameter to the wished level.  Then on a single PC one can run 
typically a N=10^5 problem in a matter of days, what would require a large cluster
at the same speed with the O(N^2) "exact" straight algorithm, or with special purpose
GRAPE hardware.   Of course nowadays the best of the best is to build a cluster of
PC's with each one or several GRAPE boards, which is not pure Beowulf since the
GRAPE boards do not belong to commodity hardware.

On the question of ring network vs. switch in your setup, for the O(N^2) algorithm most
of the time is spent in computation as soon as N exceeds a few 1000, so the 
communication method is of secondary importance.

	Dan

steve_heaton at ozemail.com.au wrote:
> G'day all
> 
> I humbly kneel before the great collective and seek their tolerance of someone new to the Dark Art of Beowulf clustering ;)
> 
> Despite having spent 14 years in a past life working on "commercial" IT&T, I have now found the true use of such power is in the support of science. I'm slogging towards my Masters in Astronomy and Large Scale Structure dynamics is a passion. To this end I've built a small cluster to do some Nbody sims (and learn a bit about Beowulf clustering).
> 
> The short description is 4x dual P3's with gigabit Ethernet to an unmanaged switch (anyone with more than 4 nodes is just showing off ;)
> 
> The Nbody code uses MPI with a "pipe" (ring) IPC structure.
> http://carol.wins.uva.nl/~spz/act/modesta/Software/src/nbody_sh1p.html
> 
> Question: Given that the MPI is performed using a ring� would there be any advantage in using a "ring" (switchless) network topology over gigabit Ethernet?
> 
> I understand that Linux would then be responsible for the routing (as opposed to the dedicated hardware in the switch). But in return, I should get deterministic performance as opposed to probabilistic. Obviously the performance bottom line is what counts regardless!
> 
> Thoughts, comments, criticisms gratefully accepted. If the idea of a ring is a dud I can save myself the bucks! :)