[MITgcm-devel] compiler optimization on SunOS

[Baylor Fox-Kemper]

Hi,
   Yes, I have found that 03 optimization at GFDL outperforms O4 and  
O5 with the xlf compiler.  Also, I found neglible improvements from  
letting the compiler compile across multiple .F file (the -ipo flag).
   As an added benefit, it doesn't take as long to compile!
   I think the real lesson is that part of setting up a set of runs  
for large-scale production is testing the compiler flags, rather than  
assuming that higher is better.  As D. and I note, it isn't clear  
that higher is faster, and as Martin has demonstrated, sometimes  
higher is even wrong!
    -Baylor

On Feb 19, 2007, at 10:00 AM, Dimitris Menemenlis wrote:

>> I guess that depends on the compiler, right?
>
> Yes, the results I quoted were for ifort on an altix using propack  
> 5 but the general lesson is probably valid for other compilers too,  
> that is, if speed matters we cannot cannot assume that higher  
> optimization levels are always faster.  Details attached for MITgcm- 
> devel reference:
>
>> I did these timings on a development system with no one else  
>> running on
>> the machine. The operating system is propack 5 that is not yet  
>> ready for
>> prime time including lack of a working CXFS. In effect I had a  
>> dedicated
>> system to run on. The optimizer has no clue about the operating  
>> system or
>> the filesystem, so I expect the results to be reproducible. That  
>> said, we
>> all know how large timing variations have been, and my version  
>> doesn't
>> include Ken's contribution to speeding up the code.
>
>
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel