This Project

The project as a whole has gone smoothly, and broadly to plan. The main problem with the project was a lack of time. When the project was proposed, the work was split into four phases with the first phase estimated to take two months, the next four, the third two and the fourth phase four months for a total of a year. Because the fourth phase was not funded, the project as a whole was shortened to eight months. Unfortunately it has been anticipated that much of the general optimisation work would be carried out during this fourth phase, since only then would the full performance implications be clear. The omission of this phase from the project meant that the optimisation of the band-scheme had to be squeezed into the other eight months, with the result that the workload was rather more intense than expected and the final result not quite as polished as had been hoped. It also meant that the third phase of the project, the band-independent optimisation algorithm, was never likely to be generally useful.

The problems with time outlined above were exacerbated by the inclusion of a `work package 0', without a corresponding extension to the overall timeline. Whilst a general analysis of Castep performance was necessary, the detailed option-by-option study requested by NAG was unexpected and it would have been better to set aside an extra month for this work rather than try to fit it into the existing timeframe.

One final frustration with the general schedule of the project was the holiday allowance. Due to various negotiations between NAG and the University of York regarding the precise financial arrangements, the holiday allowance was set at seven weeks, which was of course extremely generous. Unfortunately at nearly 20% of the project time it was completely impractical to actually use most of this allowance and still expect to get the work done.

Despite these misgivings, the project has been very successful. The original goal of the project was to improve the scaling of Castep with cores by `a factor of 8 or more to over 2000'. Despite the lack of the fourth phase's parallel band-independent optimiser, the scaling of Castep has been improved by about a factor of four, and with the extra work on $\Gamma$-point optimisations the target factor of eight or more in performance awaits only the integration of the band-parallel developments with the latest Castep source.

The considerable computational resources of HECToR, coupled with the developments in Castep detailed in this dCSE project report, should enable accurate DFT calculations on a far greater scale than has been possible to this point.