Substantial gains in performance were observed on implementing
task-farming for all three types of calculation.
The highest gains were obtained by maximising the number of workgroups
(up to the limit where the load becomes unbalanced).
This result is expected
because the ChemShell level of parallelism involves very little overhead
compared to parallel evaluation of the energy and gradient.
Single point calculations
were good predictors of the final performance and this initial step is recommended
as a method for determining the optimal number of workgroups in advance.
Tom Keal
2010-06-29