The main aim of future work will be to use the task-farming approach on HECToR for scientific applications. An ongoing collaboration with Prof. Catlow's group at UCL will use task-farmed calculations for the study of heterogeneous catalysis, with systems similar to the QM/MM test cluster that was used to benchmark the parallel NEB method. The task-farmed approach will also be useful in other areas where large-scale calculations are required, such as biomolecular modelling.
There are also still opportunities for further technical developments.
First, the task-farming method could be extended to other types of chemical calculation,
such as multiple trajectory calculations in molecular dynamics. We also plan to parallelise
the finite-difference Hessian code in DL-FIND, which is used by several optimisation
algorithms (this is equivalent to the force
command in ChemShell and will benefit
in the same way). Second, further modifications to GAMESS-UK could be made to allow
alternative parallelisation methods (ScaLAPACK, global arrays) to work in the
task-farmed environment. This is expected to be more technically challenging than
the minor modifications required for PeIGS. Third, other codes could be
modified to accept the MPI_COMM_WORKGROUP
communicator and so work in
task-farmed mode. The internal DL_POLY code in ChemShell is an obvious first candidate for this work,
but collaborations with the developers of other external software packages is also
a possibility.
The task-farming implementation will be made generally available as part of ChemShell version 3.4, which is scheduled for release in July 2010. Until then the pre-release code is available on request.
The work detailed in this report will be submitted for publication to Proceedings of the Royal Society A as part of a special issue on the activities of the Materials Chemistry Consortium.
Tom Keal 2010-06-29