bang for buck?

[Josip Loncaric]

"Geoffrey D. Jacobs" wrote:
> 
> Generally, SMP should help performance compared w/ 2 single CPUs (by
> Amdahl's rule) as average communcation speed is much higher.  All in
> all, though, it depends on the application.

The phrase "it depends" is more accurate than the word "generally."  For
a comparison of NAS benchmarks on singles/duals using different MPI
libraries and comparing TCP (fast ethernet) to VIA (Giganet cLAN)
transports, see:

  http://www.icase.edu/~josip/MPIonCoral.html

Contention for shared resources (e.g. memory) and juggling different
connections (network/internal) on duals has significant effects.  In my
tests, all 16-process NAS benchmarks run faster and more stably using
16x1 CPUs than 8x2 CPUs, regardless of the MPI library or network
transport.

The moral: Don't confuse high intra-node bandwidth with application
performance.  Other bottlenecks often dominate.

Sincerely,
Josip


-- 
Dr. Josip Loncaric, Research Fellow               mailto:josip at icase.edu
ICASE, Mail Stop 132C           PGP key at http://www.icase.edu./~josip/
NASA Langley Research Center             mailto:j.loncaric at larc.nasa.gov
Hampton, VA 23681-2199, USA    Tel. +1 757 864-2192  Fax +1 757 864-6134