Massively parallel DMRG code for quantum chemistry
Jiri Brabec1 , Jan Brandejs1 , Karol Kowalski2 , Sotiris Xantheas2 , Ors Legeza3 , Jiri Pittner1 , Libor Veis1
1J. Heyrovsky Institute of Physical Chemistry, v.v.i., Academy of Sciences of the Czech Republic,
Dolejskova 3, 18223 Prague 8, Czech Republic
2William R. Wiley Environmental Molecular Sciences Laboratory, Battelle, Pacific Northwest National Laboratory, K8-91, P.O.Box 999, Richland, Washington 99352, USA
3Strongly Correlated Systems Lendulet Research Group, Wigner Research Centre for Physics, H-1525 Budapest, Hungary
In this project we develop a massively parallel implementation of the quantum chemical density matrix renormalization group method (DMRG), which is a very powerful multireference method, however highly scalable implementations have not been developed yet. The new implementation has a potential to open the way for computations of fairly large challenging problems in quantum chemistry where strong electron correlation plays an important role, such as bio-inorganic systems with multiple transition metal centers, especially in combination with the QC methods describing properly dynamical correlation effects.
The most demanding part of the DMRG calculations is the contraction of the superblock Hamiltonian with a wave function vector. This is a sparse problem, but using the quantum symmetry sector decomposed form, sparse matrix operations are replaced by a huge number of dense matrix-matrix multiplications. We use a parallel scheme based on the MPI in combination with the OpenMP approach. The individual matrix-matrix multiplications represent elementary tasks, which are grouped into task pools distributed among nodes and executed locally. The scheduler supports dynamical balancing and an user could control the pool size in order to achieve optimal granularity on a given computing system. For handling of the large data sets in the part of the parallelization scheme we use the Global Arrays (GA) framework.