[MITgcm-support] new cluster advice

[Ed Hill]

On Mon, 16 Feb 2009 16:11:39 +0000 "Blundell J.R." wrote:
>
> Just to echo Martin's comments: CFD codes all suffer from this
> problem.


Yes!  And not just CFD!  There are plenty of problem domains (e.g.,
computational electromagnetics, image & signal processing) which tend
to be memory bandwidth limited on modern hardware.  Its been this way
for years because increases in the number and speed of cores tends
to outstrip increases in memory performance.

For some recent benchmarks, please see these SPEC URLs:

  http://www.spec.org/cpu2006/Docs/
  http://spec.org/cpu2006/results/res2009q1/

where the best recently reported floating-point (CFP2006 Base) scores
are roughly:

  Opterons:       19
  Intel Core2:    25
  Intel Core i7:  33

which is a pretty substantial increase for the new Intel Core i7
("Nehalem") chips.  If you want to dig a little deeper and look at
individual scores then I'd guess that the benchmarks labeled:

  481.wrf 437.leslie3d 459.GemsFDTD

are most likely to behave in a manner similar to MITgcm since they are
also time-stepping finite-volume or finite-difference engines.  However,
there is ultimately no substitute for running your code with *your*
inputs to see how it does...  :-)

Ed

-- 
Edward H. Hill III, PhD  |  ed at eh3.com  |  http://eh3.com/