[Mitgcm-support] RE: praise be to MITgcm

[mitgcm-support at dev.mitgcm.org]

Hi Martin/Stephan,

 An MPI parallel run requires three code and build configuration
settings:

 1. the SIZE.h thing **but** you need nSx=1, nSy=1, nPx=2, nPy=2. The
nSx and nSy settings are for multi-threaded parallelism, which is
usually more fussy than multi-process.

 2. In CPP_EEOPTIONS you need to switch to have
    #define ALLOW_USE_MPI
    #define ALWAYS_USE_MPI

 3. Use the MPI flag in genmake, as you did. However, the path names for
the mpif77 command etc... are not standardized and so can vary for
different installaitons. I would guess Stephan will know whether these
are different for you Sun install.

 Finally running needs a special command, usually called mpirun, that is
provided by the MPI system install on a given platform. This can also
vary for different installs, but Stephan will probably know about this. 

  Section 4.3 of the online manual and the subsequent text has material
on the steps above.

  At four degree resolution it is generally hard to get spectacular
speed ups through parallelism. The individual tiles become very small
and so the communication overheads show up fairly quickly. I'll be
interested to know how it goes and what you find.

Chris



> -----Original Message-----
> From: Martin Losch [mailto:mlosch at mit.edu] 
> Sent: Monday, February 24, 2003 3:41 AM
> To: cnh at plume.mit.edu
> Cc: Stephan Frickenhaus
> Subject: Re: praise be to MITgcm
> 
> 
> Chris,
> 
> PPS.
> MITgcm and MPI. We are trying, just for fun and as benchmark for our 
> expensive and worthless Sun Fire, to run the MITgcm 
> global_ocean.90x40x15 on the Sun Fire in parallel mode. What 
> do I have 
> to do to experiment, to make it work? I would do this (I have 
> absolutely 
> no experience with parallel machines, and Stephan 
> Frickenhaus, who helps 
> me has no experience with the MITgcm):
> in SIZE.h change the parameters to (to run on 4 cpu's)
>        PARAMETER (
>       &           sNx =  45,
>       &           sNy =  20,
>       &           OLx =   3,
>       &           OLy =   3,
>       &           nSx =   2,
>       &           nSy =   2,
>       &           nPx =   1,
>       &           nPy =   1,
>       &           Nx  = sNx*nSx*nPx,
>       &           Ny  = sNy*nSy*nPy,
>       &           Nr  =  15)
> 
> in genmake, I found the case SunOS+mpi, that should take care of 
> everything else shouldn't it.
> Please tell me what else I have to do.
> 
> Martin
> 
> Chris Hill wrote:
> > Martin,
> > 
> >  Did you leave us any useful netCDF material lyinbg around? 
> I am doing
> > so IO related stuff in WRAPPER2 and want to bet some netCDF 
> capabilities
> > in there if possible ( just for you :-) ). 
> > 
> >  How are the OPA setups?
> > 
> > Chris
> > 
> > 
> > 
> > 
> >>-----Original Message-----
> >>From: Chris Hill [mailto:cnh at plume.mit.edu] 
> >>Sent: Thursday, February 13, 2003 9:57 AM
> >>To: 'Martin Losch'; 'Alistair Adcroft'; 'Chris Hill'; 
> >>'Jean-Michel Campin'; 'Patrick Heimbach'
> >>Subject: RE: praise be to MITgcm
> >>
> >>
> >>Hi Martin,
> >>
> >> Great - I will put this up under our quote of the month 
> >>feature on the web site. Its probably about time to remove 
> >>the "neither Alistair or I have the slightest interest in 
> >>coupled modeling" quote that John provided a few months ago 
> >>anyhow. OK :-)..... 
> >>
> >>What sort of experiemental configurations are you trying to set up? 
> >>
> >>Can't you just change the MITgcm makefile to create an 
> >>executable called "opa.exe" and get on with it!
> >>
> >>Hope the family is well, 
> >>
> >>Chris
> >>
> >>
> >>>-----Original Message-----
> >>>From: Martin Losch [mailto:mlosch at awi-bremerhaven.de] 
> >>>Sent: Thursday, February 13, 2003 9:44 AM
> >>>To: Alistair Adcroft; Chris Hill; Jean-Michel Campin; 
> >>
> >>Patrick Heimbach
> >>
> >>>Subject: praise be to MITgcm
> >>>
> >>>
> >>>Hi guys,
> >>>
> >>>I am currently trying to learn OPA. It's a nightmare (o:
> >>>The directory structure is completely confusing and OPAque and the 
> >>>environment (wrapper) is completely broken. It took me three days, 
> >>>several start from scratch and the consulatation of various 
> >>
> >>people to 
> >>
> >>>get a simple example experiment to COMPILE on my slow Solaris 
> >>>machine. 
> >>>Now the model stops after reading the input namelist, which I 
> >>>consider a 
> >>>great success. Yesterday, purely for nostalgic reasons, I 
> >>>downloaded the 
> >>>MITgcm cd'ed into verification and typed testscript, just to see 
> >>>SOMETHING work, although it didn't make me feel better, that 
> >>>everything 
> >>>went smoothly as I expected ...
> >>>
> >>>Martin
> >>>-- 
> >>>Martin Losch
> >>>Alfred Wegener Institute for Polar and Marine Research
> >>>Postfach 120161, 27515 Bremerhaven, Germany;
> >>>Tel./Fax: ++49(0471)4831-1872/1797
> >>>
> >>>
> >>
> > 
> 
> 
> -- 
> Martin Losch
> Alfred Wegener Institute for Polar and Marine Research
> Postfach 120161, 27515 Bremerhaven, Germany;
> Tel./Fax: ++49(0471)4831-1872/1797
> 
>