[MITgcm-devel] going to update build option for pleiades

[Menemenlis, Dimitris (3248)]

David, thank you for checking build-option file.

Hong, please go ahead and commit new file to CVS.

The one item that I like about this new build-option file
is -axSSE4.2,SSSE3 which optimizes automatically for
the different pleiades CPUs and then uses the correct
optimized code at runtime.

Cheers

Dimitris Menemenlis

On Jun 16, 2011, at 8:55 AM, David Ferreira wrote:

> Hi Hong,
> I did a test with your new optfile (in attachment). The June 15 test 
> were done with the old file and the June 16 were done with yours (same 
> code). There are just a couple of minor differences.
> I haven't put the results of the restart test because they are identical 
> with the 2 opt files.
> Looks all good to me.
> david
> 
> 
> On 6/14/11 1:21 PM, Hong Zhang wrote:
>> Dear all,
>> Recently pleiades has many upgrades (OS and modules),
>> so we're going to change the build option accordingly.
>> some suggested change:
>> 1. use new modules of compiler/mpi/netcdf and change associated 
>> INCLUDE/LIB
>>   (module load comp-intel/2011.2 mpi-sgi/mpt.2.04-fsa netcdf/4.0)
>> 2. FFLAGS: remove -w95 (deprecated soon)
>> 3. FOPTIM:. add -axSSE4.2,SSSE3, add -traceback, remove -xSSE4.2
>> some justification from Art:
>> ---
>> 
>> for SLES11 OS, users have to use mpt.2.04. The earlier MPT modules 
>> won't work
>> reliably on SLES11;
>> The module mpi-sgi/mpt.2.04-fsa differs from mpi-sgi/mpt.2.04 in that it
>> attempts to use "multi-rail" networking to improve message passing 
>> throughput;
>> There are also some optimizer/vectorizer improvements in v12 compilers 
>> (comp-intel/2011.2);
>> -axSSE4.2,SSSE3 instructs
>> the compiler to build two execution paths in an executable, one for each
>> architecture. At run time, the executable checks the CPU it's running
>> on, and then runs the optimized choice. The above produces code 
>> optimized for both Harpertown and Nehalem/Westmere instruction sets;
>> Since the use of -traceback produces much more meaningful runtime info
>> at no cost, I recommend always including the option.
>> ---
>> We've tested above change and it's OK.
>> Any comments is welcome before check-in.
>> hong