Need help & comments on HPL benchmark

Hello all,

I've been running HPL tests on a 8 node Linux cluster. Each node has 512MB
DDR RAM and Intel Pentium IV 2.0 GHz CPU. I have a few questions.

What is the "theoretical peak" value for this cluster? According to Jack
Dongarra's list (August 2003), an Intel Pentium IV 2.0 GHz machine has a 2 x
2000 MHz = 4 Gflops theoretical peak. So for 8 nodes that means 32 Gflops in
total. Is that right?

It is said that, normally, one should expect to reach approximately 8.5% of
the "theoretical peak" on a fast ethernet based cluster. So that means
approx. 2.72 Gflops for our cluster. The highest value I've been able to get
(when I use VSIPL library with HPL) is 1.35 Gflops (N=5000). For N>5000 the
value drops. But if I use ATLAS lib., which is optimized for Pentium IVs and
SSE2, I get 4.4 Gflops (N=10000). This seems weird because it is more than
the theoretical peak value. So it probably means that one of our assumptions
was wrong. What do you think about it? Should the theoretical peak for PIVs
be calculated differently? Is that the SSE2 which changes the results? Or
can a cluster based on fast ethernet, reach more that 8.5% of the T.P.?

I need all your comments.

Thanks all in advance.