[MITgcm-support] MITgcm on AS4, kernel 2.6

Chris Hill cnh at mit.edu
Thu Oct 6 17:49:17 EDT 2005


Hi Hezi,

  Is there any chance it could be a hardware or OS problem?

Chris

Ed Hill wrote:
> On Thu, 2005-10-06 at 18:54 +0300, Hezi Gildor wrote:
> 
>>Hi Ed,
>>
>>thanks for the prompt reply.
>>
>>The OS was AS4.0 and AS4U1. The compilers tried are gcc and pgi (5.2 and 
>>6.0). Tried using with LAM, MPICH (from pgi as well as built from tar 
>>ball), MPICH2. The compiler options used were many. From the default 
>>options to the -O3, to -fast, -fastsse, and -kieee (last 3 for pgi).
>>
>>Communications is through the internet. (No private network for now).
>>
>>Dual-opteron machines with
>>
>>AMD Opteron(tm) Processor 250
>>
>>The code is based on MITgcm_ss_20050719, checkpoint571_post. We have few 
>>hard-wired changes to code but it runs ok on Alpha machine and on 
>>cluster with Linux Kernel 2.4 so i don't think that the problem is with 
>>the code.
>>
>>when we compared fields such as salinity between two runs with same 
>>executables, we can get that at most grid points the difference is zero 
>>but at many isolated points scattered throughout the domain it is not (O 
>>of 10^-6).
> 
> 
> Hi Hezi,
> 
> If you are using *exactly* the same executable and *exactly* the same
> inputs then, assuming you are running it on exactly the same machine, I
> have no idea why you're not seeing exactly the same answer.  Of course,
> if anything changes from one run to another (eg. different compiler,
> different OS, different hardware, etc.) then you won't, in general, see
> exactly identical ("zero difference") output due to differences in
> numerical roundoff, etc.
> 
> And, to be fair, your description is not enough information for us to
> have a good idea of what you're doing (or trying to do).  So, can you
> narrow this problem down to a (hopefully simple and) repeatable example?
> That is, can you create a small but complete setup that includes *all*
> the code changes that you've made and *all* your input files?
> 
> Once you can create an example that repeatably demonstrates a problem,
> then we can work with you to try to determine the reason.  But, without
> such an example, I'm afraid that all we can offer you is guesses...
> 
> Ed
> 




More information about the MITgcm-support mailing list