Dear Gary,
It's finally finished! After almost 2 years, I have finally made the
multi_processor_merge branch 100% functional again!! Your brilliant
multi-processor framework appears to be working at full capacity with
all the advances of the relax 1.3 line. The parallelised diffusion
tensor grid search is activated again, though I have made a completely
new implementation which relies on a few new types of non-linear grid
searches added to the minfx project (https://gna.org/projects/minfx/).
You can use minfx without installing it in relax by checking out a
directory of the main line 'trunk' in the base relax directory:
svn co http://svn.gna.org/svn/minfx/trunk/minfx
I have made many code changes, code clean ups, and fixed many bugs.
There has been a lot of work with STDOUT and STDERR: interleaving is
now mostly correct, and I have tamed the loss of control of STDOUT and
STDERR problem - hence the test suite in the uni-processor fabric is
now functional. The only issue I could not fix was the STDOUT and
STDERR prefixing in mpi4py mode, but this is one we have discussed
before that has no solution. Uni-processor mode is seamless and you
cannot even tell that is it passing through the multi-processor
framework - an essential requirement for merging this back. As I have
dual-core machines and openmpi installed, I have tested the mpi4py
fabric with:
mpirun -np 6 ./relax --multi mpi4py
test_suite/system_tests/scripts/omp_model_free.py
and this completes perfectly. I was wondering if you could quickly
test this again on your massive beowolf cluster? It would be good to
see if I haven't broken anything and that the scaling as shown in
https://mail.gna.org/public/relax-devel/2007-04/msg00048.html is still
there.
I would also like to terminate this branch - i.e. finally merge this
beast into the 1.3 line. For this, I would like to discuss with you -
in private might be best - how you wish to publish this and how this
should be synchronised with a new relax release introducing the
multi-processor framework. If you don't have any system you applied
this to for a model-free analysis, I think this would by itself make a
nice little communication, just announcing to the world its existence.
Or a few published data sets could be very quickly reanalysed. I'm
looking forward to hearing from you.
Cheers,
Edward
Completion of the multi_processor_merge branch.