[MITgcm-support] mpi with 4 processors
Holly Dail
hdail at MIT.EDU
Mon Nov 1 13:56:12 EDT 2010
Hi Angela -
Your problem, especially the 2-D one, may simply be too small to benefit from more than 2 processors (i.e. the added communication costs of using 4 isn't compensating for the faster computation). Have you found that the nSx=6 is helpful for your machine? I'm currently running on 24 processors (Pleiades) with the following SIZE.h:
& sNx = 20,
& sNy = 27,
& OLx = 3,
& OLy = 3,
& nSx = 1,
& nSy = 1,
& nPx = 6,
& nPy = 4,
& Nx = sNx*nSx*nPx,
& Ny = sNy*nSy*nPy,
& Nr = 23)
Holly
On Nov 1, 2010, at Nov 1 , 1:44 PM, Angela Marie Zalucha wrote:
> Hi,
>
> I am trying to figure out how to use the MIT GCM with multiple processors. I found the examples for how to modify SIZE.h for 2 processors. In 2D I use:
>
> sNx = 1,
> sNy = 32,
> nSx = 1,
> nSy = 1,
> nPx = 1,
> nPy = 2,
>
> In 3D I use:
>
> sNx = 16,
> sNy = 32,
> nSx = 6,
> nSy = 1,
> nPx = 2,
> nPy = 1,
>
> Indeed, there is a significant improvement in speed over 1 processor. I have up to 4 processors available to me, but the configurations I have tried with 4 processors don't run any faster than with 2. Is there any optimum configuration I can use with 4 processors? Or 3 processors if 4 is not possible?
>
> It might be that I'm not understanding how to scale things for more processors, so any example with n > 2 would be helpful.
>
> I use gfortran for the compiler and mpich2 for mpi. I am running on computers at Southwest Research Institute in Boulder, CO. I am using 30 vertical levels.
>
> Angela
>
> _-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
> Angela Zalucha, Ph.D.
> Postdoctoral Associate
> Department of Earth, Atmospheric,
> and Planetary Sciences
> Massachusetts Institute of Technology
>
> Office located at Southwest Research Institute
> 1050 Walnut Street, Suite 300
> Boulder, CO 80302
> (720) 208-7211
>
>
> http://web.mit.edu/azalucha/www/index.html
> _-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
>
>
>
>
> _______________________________________________
> MITgcm-support mailing list
> MITgcm-support at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-support
More information about the MITgcm-support
mailing list