COSINUS - Conception et Simulation

Wavelets and order N for very large scale calculations of Electronic structure – NEWCASTLE

Submission summary

Ab initio methods are well used in the fields of solid state physics and quantum chemistry because they are free of parameter. They solve the Schrödinger equation in the framework of the density Functional Theory.
Traditionally, these methods use gaussian or plane wave basis sets which limit the scalability. Moreover the CPU time varies as a power of 3 of the number of atoms.

With the software BigDFT, we have shown that it is feasible to use real space basis sets as wavelets with the same accuracy as traditional basis sets. Due to adaptive properties, the memory consumption can be reduced drastically compared to plane wave basis set. BigDFT has also a better scalability. We have demonstrated an efficiency over 85% up to 1000 cores. Moreover, order N method which scale linearly with the number of atoms can be used with real space basis sets.

The first goal of this project is to develop an ab initio method based on wavelet which can exploit petascaling supercomputers by means of order N methods and a better memory consumption using softer pseudopotentials.
Hundred of thousands cores could be used to solve multiscale problems as involved in biology or new devices as OLED.
Moreover all developed components could be re-used in other softwares (ABINIT, CP2K, Octopus, ...) as it is already done with present versions of BigDFT (Poisson solver and exact exchange calculations).

The second goal is to have a robust and reliable Message Passing Interface library based on OpenMPI with fault tolerance. In order to deliver a test suite based on BigDFT for the future petascaling supercomputer (PRACE project), we had tested the supercomputers of national centers (CCRT and CINES) up to 600 cores. We found that the collective communications had not an efficient scalability. Moreover, all MPI libraries had some trouble as the number of cores increases which were not very reproducible. We think that it is very important to have a test suite to optimize algorithms for collective communications versus the number of cores and that fault tolerance is a very crucial concept with the arrival of petascaling computer.

With this project, we would have a reliable ab initio methods adapted for large scale simulation, a robust message passing interface with fault tolerance for massively parallel computers and a test suite to optimize algorithms of collective communications. All developments will be available and re-usable for other softwares.

Project coordination

Thierry DEUTSCH (COMMISSARIAT A L'ENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVES ET AUX ENERGIES ALTERNATIVES) – thierry.deutsch@cea.fr

The author of this summary is the project coordinator, who is responsible for the content of this summary. The ANR declines any responsibility as for its contents.

Partner

INAC/SP2M COMMISSARIAT A L'ENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVES ET AUX ENERGIES ALTERNATIVES
ESRF EUROPEAN SYNCHROTRON RADIATION FACILITY - ESRF
CEA-DIF-DPTA COMMISSARIAT A L'ENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVES - Direction des Applications Militaires
BULL BULL SAS

Help of the ANR 660,088 euros
Beginning and duration of the scientific project: - 36 Months

Useful links

Explorez notre base de projets financés

 

 

ANR makes available its datasets on funded projects, click here to find more.

Sign up for the latest news:
Subscribe to our newsletter