[Beowulf] hpl size problems

[Mark Hahn]

> Our HPL run yielded 1516 Gflops and 83.4% efficiency. According to Dell,
> this broke the efficiency records for a system of this configuration

interesting.  HPL is pretty simple, computationally, so it seems to me
that if you put increasing CPU power in each node, and minimize the 
amount of communication necessary by also increasing the work/node (memory),
then even with a fixed interconnect, you should expect to gain efficiency.

incidentally, shouldn't modern x86 be rated at more than 2 flops/cycle?

> (we like to think Warewulf had something to do with that ;).

also interesting, why?

> hours) running on Centos-3.5 and saw a pretty amazing speedup of the
> scientific code (*over* 30% faster runtimes) then with the previous
> RedHat/Rocks build. Warewulf also makes the cluster rather trivial to

such a speedup is indeed impressive; what changed?

> We did find that symbol errors in the fabric are very common if anyone
> "breathes" on the wire plant and cause drastic changes in performance.

hmm, interesting.  I guess that sensitivity would also apply to other 
HSI that uses the same phy layers (infinipath, rapidarray).  does anyone
have a myri-10G cluster they can comment on?  (I'm happy with the robustness
of my myri-2g and also with my older quadrics.)

regards, mark hahn.