[Beowulf] Re: building a new cluster

[Joachim Worringen]

SC Huang wrote:
> 5. I only have wall-clocking timing of my code on various platforms. I 
> don't know how sensitive it is to cache size. I guess the bigger cache, 
> the better, because the code is operating large arrays all the time.

If your arrays are really large, than it's the memory bandwidth that 
counts, not the cache size. Or your ability to implement cache-friendly 
(blocked) versions of your code (probably with the assistance of 
libraries like ATLAS).

Just a side note on this: we have a 32-node dual-Athlon cluster with 
Myrinet in the lab as well as an SX-6i (single processor SX-6). We've 
seen our financial modelling codes run faster on the SX-6i than on the 
cluster (with a non-trivially parallized version). And in contrast to 
the cluster, the SX-6i is running non-stop for months (as long as we 
have it here). It's much smaller and more energy-efficient, too - but 
it's not cheaper.

  Joachim

-- 
Joachim Worringen - NEC C&C research lab St.Augustin
fon +49-2241-9252.20 - fax .99 - http://www.ccrl-nece.de