[MITgcm-support] IO problems with large numbers of CPUs

Matthew Mazloff mmazloff at MIT.EDU
Sun Jul 13 09:48:51 EDT 2008


Hi Christopher,

im not sure how globalfiles works, but you can overload the IO  
processor with SingleCpuIo if you have
#define DISABLE_MPI_READY_TO_RECEIVE
...so try with that off.

And you can reduce the memory size the diagnostics package uses by  
changing
PARAMETER( numdiags = 7*Nr )
in DIAGNOSTICS_SIZE.h
...reduce to the minimum you need

sorry i can't be of more help

-matt





On Jul 11, 2008, at 3:39 PM, Dr. Christopher L. Wolfe wrote:

>
> Hi All,
>
> I'm trying to get the MITgcm to run on 1024 CPUs on a Cray XT4  
> (franklin.nersc.gov). Per-tile IO works fine, but produces an  
> utterly unmanageable number of files. However, if I turn on  
> SingleCpuIo or GlobalFiles *and* the diagnostics package, my jobs  
> die with an out of memory error. SingleCpuIo or GlobalFiles works  
> fine without the diagnostics package and the diagnostics package  
> works fine without SingleCpuIo or GlobalFiles. Changing these  
> options shouldn't have that much of an impact on memory usage,  
> since nearly all MITgcm's memory is statically allocated. I figure  
> that SingleCpuIo might be overflowing my MPI message buffers, but I  
> can't see why GlobalFiles would if each process writes to the  
> global file independently. If anyone has any idea what's going on  
> here, I'd love to hear your input.
>
> Thanks,
> Christopher
>
> -----------------------------------------------------------
> Dr. Christopher L. Wolfe              	   858-534-4560
> Physical Oceanography Research Division    OAR 357
> Scripps Institution of Oceanography, UCSD  clwolfe at ucsd.edu
> -----------------------------------------------------------
>
>
>
> _______________________________________________
> MITgcm-support mailing list
> MITgcm-support at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-support




More information about the MITgcm-support mailing list