[Mitgcm-support] First and PRELIMINARY findings with new code
mitgcm-support at dev.mitgcm.org
mitgcm-support at dev.mitgcm.org
Wed Jul 9 15:47:01 EDT 2003
Carl, John, Alistair and Chris;
Patrick is on its way back to the east coast and he will report more on this
matter once he is back. We made progress here which allowed us not to start
parallel experiments with both codes. Quite obviously there are still interfaces
missing related to the optimization.
But the important thing is that we started to compare timings, storage
requirements, etc between the old and the new code. In some sense the numbers we
have so fare are preliminary and we will continue along those lines. However,
the findings we have by now are somewhat alarming and frustrating. Our numbers
are still preliminary but I think the following is worth some warning:
1) The old code runs on the Origing 2000 with automatic parallelization about as
efficient as the new code - this is true on 12 CPUs and for a 2 degree model.
Results could look different with more CPUs and/or higher resolution.
2) The new code requires at least 5 times more core memory through a
substantially higher storage requirement. Patrick suggested that he can reduce
that by some fraction. But the primary differences could reside in the different
time stepping schemes that create different dependencies. I will ask Ralf to
check carefully all the store directives implemented right now. Several might
not be required as a recomputation could be much more efficient - if it is
small. Pending further confirmation - as the code stands right now we could not
do the computations I am doing with the classic code since we could not effort
the storage.
For more details I refer to Patrick's verbal report.
Detlef.
--
Detlef Stammer
Physical Oceanography Research Division
Scripps Institution of Oceanography
9500 Gilman Dr.; MS: 0230
La Jolla, CA 92093-0230
ph: (858) 822-3376
fax: (858) 534-9820
email: dstammer at ucsd.edu
For FeDex Delivery:
8605 La Jolla Shores Dr.
La Jolla, 92037.
More information about the MITgcm-support
mailing list