Conclusions

The dCSE project ''Improving the parallelisation and adding functionality to the quantum Monte Carlo code CASINO'' has achieved its stated goals. CASINO is now capable to compute more complex physical systems with increased efficiency on the latest generation of supercomputers.

We summarise the main improvements implemented in the new version of the code by the dCSE work:

• The System V shared memory solution allows sharing of the OPO data and therefore each core of a processor can run a computing task even for very large scale models. In quantitative terms a computation that needs from 2 to 4 GB of OPO data increases its speed with a factor of 2, with respect to the initial version of the code, and with a factor of 4 if the size the OPO data are larger that 4GB (for quadcore processors).

• The input optimisations have eliminated unnecessary waiting time for data input which ranged from 30 to 60 minutes saving in this way from 3000 to 6000 AU per 1000 tasks run.

• The computation using mixed mode second level parallelism increases the computation speed with a factor between 1.8-1.6 for systems with a more than 1000 electrons using quadcore processors. The mixed mode parallelism offers the possibility to perform calculations for very large systems (about 10,000 electrons).

The System V shared memory algorithm and the input optimisation have been ported in the release 2.4 of CASINO (July 2009). The CASINO manual was updated with the necessary information and the Makefile was adapted in order to help users to easily generate the executable that fits their requirements. The second level parallelism is planed to be available in the 2010 release, meanwhile it can be obtained upon request from development branches.

An illustration of the performance level reached by CASINO over the last two years can be seen in Fig which shows that good scaling was obtained for runs up to 40,000 tasks.

The degradation of scaling for very large number of tasks hints for the need to improve the configuration redistribution algorithm. Another extension which is planned in the next dCSE project is the coupling of CASINO with molecular dynamics codes. This will allow a much faster computation of the QMC corrections to the system energy using the correlated samples technique.