[ase-users] Turbomole NEB calculation in parallel

[Max Hoffmann]

Hi Leo,

    I am not familiar with Turbomole. But my take on NEB
parallelization is the following: every so often someone starts with
doing NEB calculations and wonders whether one could parallelize over
images since during each energy and force calculation there is no
communication between images. Sounds great except for the fact that
with most calculators the time required to reach convergence varies by
about 30%-50% percent depending on the geometry. Therefore if one
image takes longer all the other images (and their CPUs) simply idle
which is not good parallelization at all. I therefore tend to do it
parallel and try to parallelize each electronic structure calculation
over as many nodes as possible. If the full NEB convergence does not
fit into a cluster walltime it is not so hard to restart it using the
NEB builtin trajectory and pickle mechanism. Does that make sense?


IHTH
Max.