[MITgcm-devel] [MITgcm-cvs] MITgcm/tools/build_options CVS Commit

[Menemenlis, Dimitris (3248)]

David, I have not timed the different SSE options but Art thinks they should not make much difference.  Was the speedup that you observed due to SSE version or was it due to fact that SSE4.2-compatible nodes are newer and faster?

In terms of how cryptic the message is, neither Hong, nor Pierre, nor NAS support, nor I had a clue, initially, why it happened and how to fix. Pierre and I ended up wasting a day because of this minor issue.  Specifically, this morning instead of looking at model results we spent time debugging an optfile.

Your argument is similar to Constantinos' reasoning regarding optfile for greenplanet, that is, let's make it hard for new MITgcm users to get code up and running so that they are forced to dig in and 
get their hands dirty.  If this is the objective then we should leave optfile as is.  If objective is to provide as smooth an initial experience as possible, then we need to change to SSE4.1, with comments in optfile that 4.2 and 4.3 are available and possibly faster on some of the nodes.

Dimitris

On Apr 9, 2010, at 2:42 PM, David Ferreira wrote:

> Dimitri,
> My reasoning was more or less the opposite of yours: with -xSSE4.2, if 
> someone
> makes the mistake of submitting to queue "normal", he will get an error 
> message
> and realize there is an issue. BTW, the message is not really criptic, 
> (+documented on
> the NAS webpage+comments in the optfile):
> 
> Fatal Error: This program was not built to run on the processor
> in your system. The allowed processors are: Intel(R) processors
> with SSE4.2 and POPCNT instructions support.
> 
> With xSSE4.1 by default, most people would not realize they could benefit
> from running on normal_N with xSSE4.2 (and in my case I noticed a 
> significant
> speed up, I can't remember exactly how much though !).
> Is it not the case for you ?
> david
> 
> 
> 
> On 4/9/10 2:07 PM, Art Lazanoff wrote:
>> a few additional notes:
>> 
>> ifort v11.1 supports -xSSE4.1 and -xHost but not earlier compilers.
>> 
>> Use of -xHost is interesting if and only the code is to be run on the 
>> same processor family (harpertown or nehalem, etc) as the code is 
>> compiled on. -xHost retrieves the CPUs feature set if possible. The 
>> Pleiades front-ends are currently Harpertown. One solution is to run a 
>> PBS job to compile the code.
>> 
>> At the moment, NAS PBS offers Harpertown queues (-q normal) and a 
>> Nehalem queue (-q normal_N). The next assortment of nodes is due to 
>> arrive in a few weeks and these are Westmeres. We'll likely have 
>> another queue forthese.
>> 
>> So, Dimitris, the answer to your question from earlier today about 
>> which CPUs you ran on can be determine by which queue you submitted 
>> the job to.
>> 
>> Sincerely,
>> 
>> Art
>> 
>> ***************************************************************************** 
>> 
>> Art Lazanoff                                     NASA Ames Research 
>> Center
>> Senior Scientific Consultant                       Mail Stop 258-6
>> Arthur.S.Lazanoff at nasa.gov (650) 604-1687        Moffett Field, CA 
>> 94035-1000
>> ***************************************************************************** 
>> 
>> 
>> --------------------------------------------------
>> From: "Menemenlis, Dimitris (3248)" <Dimitris.Menemenlis at jpl.nasa.gov>
>> Sent: Friday, April 09, 2010 10:51 AM
>> To: <mitgcm-devel at mitgcm.org>
>> Cc: "Lazanoff, Arthur S. (ARC-AFS)[COMPUTER SCIENCES CORPORATION]" 
>> <arthur.s.lazanoff at nasa.gov>
>> Subject: [MITgcm-devel] [MITgcm-cvs] MITgcm/tools/build_options 
>> CVS     Commit
>> 
>>> David, can I make SSE4.1 the default in linux_ia64_ifort+mpi_ice_nas
>>> with a comment that 4.2 and 4.3 is also available on some of the 
>>> newer processors.
>>> 
>>> 4.1 is the lowest common denominator.  That way new pleiades users,
>>> e.g., Pierre Rampal, don't end up with cryptic error messages when
>>> they try to use the default optfile.
>>> 
>>> Also, according to Art Lazanoff, there is probably not much speed
>>> up to be expected from 4.2 and 4.3 over 4.1.
>>> 
>>> Let me know, Dimitris
>>> 
>>> 
>> 
>> _______________________________________________
>> MITgcm-devel mailing list
>> MITgcm-devel at mitgcm.org
>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
>