Re: ccm3 using both SMP and MPI

Subject: Re: ccm3 using both SMP and MPI
From: Norm Wood (norm@atmos.colostate.edu)
Date: Wed Oct 31 2001 - 11:15:23 MST

Thanks, Jim and others who replied directly.

Just as a followup, the suggestion to stick with SPMD mode worked fine.
I set up the mpich library to use the standard ch_p4 device, but
configured it with the shared memory option, then modified mpi's
list of machine names to show that the node had dual processors.
The model compiled and ran, using two processes on the local machine,
with no problems.

I'm also trying the reduced-resolution T5 run with threading turned on
via the compiler flag, but there seems to be a snag in the interpic utility.

In lines 69 and 75 of fmain.f90, a "read' statement is used to read an
integer from a string. The syntax used is:

   read(nx,'(i4)') arg

where "arg" is the string and "nx" is the integer variable into which the
value should be read. Should "arg" and "nx" be reversed? Otherwise,
the read statement thinks "nx" is a file descriptor and tries to read
from it and put the value into "arg".

Regards,

Norm

On 29 Oct., Jim Rosinski wrote:
> On Mon, 29 Oct 2001, Norm Wood wrote:
>
> > I'm interested in finding out if anyone has had success running
> > ccm3.10.11.brnchT.366physics.4 on dual-processor linux boxes using
> > Portland Group's PGF90 compiler.
>
> The symptoms you describe sound like the slave stacksize isn't big enough.
> The example rundeck provided with the distribution (run.pc.linux.csh) should
> set this number to a value adequate for the default model (look for setenv
> MPSTKZ 40M). To test if this is the problem, I'd try running at very low
> resolution (e.g. T5) and see if it works with threading turned on (-mp arg to
> pgf90). If you're adding constituents, history variables, or running at a
> higher resolution the 40M may not be big enough.
>
> I wish these blasted systems would just default magic numbers like MPSTKZ to
> "as big as possible", but none of them do.
>
> If the above suggestions fail, try running in SPMD mode. This allows the
> code to use multiple processors, and doesn't require the -mp option.
>
> Regards,
>
> Jim Rosinski
> CCM Core Group 

-- 
Norm Wood

     Department of Atmospheric Science
     Colorado State University
     norm@atmos.colostate.edu
     phone:  970-491-3334

This archive was generated by hypermail 2b27 : Tue Jan 22 2002 - 11:16:19 MST