John:

Just want to bounce this off you.The same tiegcm code 
that uses shared-mem OpenMP on the Origin and J90's 
(its pretty fast on ute) does not work on prospect.
It runs fine single threaded, but w/ > 1 processor, 
only the master proc gets it right. This is the *same* 
source that runs on ute (w/ a few caviats for i/o, 
libs, etc). No mpi here, only OpenMP.

Its as if some data that's meant to be global across
threads (probably regular common) and is set by the 
master proc before the parallel loops, is not getting 
to the slaves. As if it were considered local to the
master. Again, this works fine on the SGI. This is a 1-d
decomposition over latitude, where nlat=36. On the Compaq,
if I use 4 threads, only the 1st 9 latitudes (36/4) are
correct, i.e., the master's share of the decomposition.
If I run w/ 2 threads, half of them are correct. 

I make extensive use of !$THREAD_PRIVATE common, and
I know IBM has been late implementing this, but I have 
run small test codes on prospect testing both regular 
and private common and it appears to work. In the model
it looks like something is considered master-private 
when it should be global. Maybe there are still some
data statements initializing common outside block data,
but I don't think so because the IBM compiler complains
about that and I've put them in block data. Thanks for
any suggestions..

--Ben