The dCSE project

The aim of this distributed Computational Science and Engineering (dCSE) project is to add a task-farm parallel framework to ChemShell in order to parallelise tasks that are commonly employed in materials cluster simulations. The dCSE proposal was written by C. Richard A. Catlow, Alexey Sokol and Fedor Goumans (University College London) and Paul Sherwood, Huub van Dam and Johannes Kästner (STFC Daresbury Laboratory). The project work was carried out by Thomas Keal (STFC Daresbury Laboratory).

9 months of development effort were deployed over an 18 month period from January 2009 to June 2010. The code development objectives for the project were:

• ChemShell (and associated Tcl libraries) adapted to support task farmed parallelism using split communicators. Commands can be executed on specific workgroups by using conditional tests within the Tcl interpreter.
• Finite difference Hessian code within ChemShell re-written to allow task-farmed execution. Speed-up factor of 2 on a test calculation (as approved by PI) when comparing existing implementation (a single parallel task running across a large HECToR partition of over 1000 processors) with the task farmed version.
• Interface with the parallel implementation of DL-FIND such that the separate tasks within ChemShell are given independent work to perform. Demonstration of parallel NEB for transition-state determination in a surface chemistry problem, speed-up (existing relative to running the points sequentially on a large parallel platform) of 2.
• DL-FIND parallel optimisation capabilities interfaced to task farmed version of ChemShell. Demonstration of global minimum search using stochastic search and genetic algorithms. The stochastic search can also be run using the baseline code (the non-task farmed version of ChemShell) code and will show speed-up of 3.

All the objectives were achieved, and the benchmark targets surpassed, as detailed below.

The benchmark calculations in this report were performed on the XT4 component of HECToR in its Phase 2a configuration. This is a Cray machine with 2.3 GHz quad-core AMD opteron processors and 2 GB of memory per core. Due to the simple nature of the parallelisation method employed, we expect the new framework to run equally efficiently on the recently-introduced Phase 2b configuration of HECToR and on other parallel platforms. The test cases have been chosen to represent calculations that might realistically be run by users on HECToR and so should give a good indication of the performance gains expected from the new approach.

ChemShell, GAMESS-UK and GULP were compiled on HECToR using the Cray compiler wrapper ftn with the PGI C and Fortran compilers (v9.0-4).