[MITgcm-support] FW: exf_interp_read/singleCPUIO - correction

Martin Losch Martin.Losch at awi.de
Fri Dec 17 03:09:19 EST 2010

Hi Clare,

is the problem you are having related to Karen Assmann's problem (I am guessing, because both of you are at BAS an probably use the same architecture/cluster)? It does sound likely.

Does your problem only occur with exf_interp_read? 
MPI_BCAST is only used in this routine (and ini_procs, but there it's a much smaller integer array itemp(1:2) that is broadcasted).
If so, maybe it  makes sense to test the MPI_BCAST in a simple program? Maybe the MPI-implementation is fishy?
It also could make sense to test, what "ierr" returns.

As a hack (I am no expert on this) you could try to circumnavigate the useSingleCPUio business for this file and make all cpus read the same field. I'd try
1. remove       IF ( MASTER_CPU_IO(myThid) ) THEN/ENDIF statements (so that all processes read, better: replace by _BEGIN_MASTER (my_thid)/_END_MASTER( myThid ))
2. remove the useSingleCPUio blocks
This will make reading slower, especially on an NFS file system, but all process have the forcing data availabe.


On Dec 16, 2010, at 6:19 PM, Enright, Clare wrote:

More information about the MITgcm-support mailing list