[Mitgcm-support] Re: [Fwd: Re: configuring the MITgcm]

[mitgcm-support at dev.mitgcm.org]

Hi Amy,

 Alistair was feeling busy so he forwarded to me.
 I'm v. busy this week (always am actually!). Quick glance in 
 your mail, the nSy = 8 looks odd. There is an unreleased web 
 site with more info. that is close to be finished

 http://mitgcm.org/sealion2

 Take a look at 
 online documents
   using the wrapper ( section 4 )
    specifying a domain decomposition ( sub sec 4.3 )

 Can look more next week. The doc. site above 
 will be public soon, but for now don't tell anyone I
 told you where it was! 

 Sonya Legg and Mike Spall at WHOI and a whole crowd
 of people at JPL run the code on SGI's - emails
 are on mitgcm.org main contacts page.

 For 8 procs. I would stick to MPI, no threads. 

Chris

Alistair Adcroft wrote:
> 
> Chris,
> 
> Your domain....?
> 
> A.
> 
> -------- Original Message --------
> From: "amy" <amy18 at sprynet.com>
> Subject: Re: configuring the MITgcm
> To: "Alistair Adcroft" <adcroft at mit.edu>
> CC: "Amy Solomon" <amy at musashimaru.soest.hawaii.edu>
> 
> Hi Alistair,
> 
> Yes, I have been following the work that you have been doing on the web.
> It really looks great.  I have already started using your subroutine and
> parameter index.
> 
> I downloaded the man pages last week and figured out how to run
> the model on multiple processors. On an SGI Origin 2000 it takes about a
> minute per
> day to run the model for nPy>7. That seems really slow, doesn't it?
> I changed:
> nPy=8
> sNy=8
> define ALLOW_USE_MPI
> define ALWAYS_USE_MPI.
> What is the difference between those two options?
> 
> Most of the time is spent in the cg2d subroutine (I figured out that the
> zonal filter doesn't work when  nPx>1 so I have only been varying nPy). The
> documentation says that
> this routine runs faster when you use interleaving threads. I figured out
> how to get more than
> one thread running but the model won't run when I use this option.
> 
> The atmospheric physics documentation says that a T21 model configuration
> takes 1 sec
> to run 1 model day on a C90. What do you think the best model performance on
> an O2000 would be?
> 
> Thank you for your help, Alistair!
> Amy