TY - CHAP

T1 - Advances in the Parallelisation of Software for Quantum Chemistry Applications

AU - Roderus, Martin

AU - Matveev, Alexei

AU - Bungartz, Hans Joachim

AU - Rösch, Notker

N1 - Publisher Copyright:
© 2013, Springer-Verlag Berlin Heidelberg.

PY - 2013

Y1 - 2013

N2 - Density functional theory (DFT) provides some of the most important methods used in computational theory today. They allow one to determine the electronic structure of finite chemical systems, be they molecules or clusters, using a quantum-mechanical model, and exposes, thus, the great majority of the systems’ properties relevant to chemical applications. However, the numerical treatment of large chemical systems proves to be expensive, requiring elaborate parallelisation strategies.This paper presents two recent developments which aim at improving the parallel scalability of the quantum chemistry code ParaGauss. First, we introduce a new Fortran interface to parallel matrix algebra and its library implementation. This interface specifies a set of distributed data objects, combined with a set of linear algebra operators. Thus, complicated algebraic expressions can be expressed efficiently in pseudo-mathematical notation, while the numerical computations are carried out by back-end parallel routines. This technique is evaluated on relativistic transformations, as implemented in ParaGauss.The second development addresses the solution of the generalized matrix eigenvalue problem—an inherent step in electronic structure calculations. In the case the symmetry of a molecule is exploited, pertinent matrices expose a block-diagonal structure which makes the efficient use of existing parallel eigenvalue solvers difficult. We discuss a technique that uses a malleable parallel task scheduling (MPTS) algorithm for scheduling instances of parallel ScaLAPACK-routines on the available processor resources. This technique significantly improves the parallel performance of this numerical step, reducing the corresponding execution time to below 1 s in most applications considered.

AB - Density functional theory (DFT) provides some of the most important methods used in computational theory today. They allow one to determine the electronic structure of finite chemical systems, be they molecules or clusters, using a quantum-mechanical model, and exposes, thus, the great majority of the systems’ properties relevant to chemical applications. However, the numerical treatment of large chemical systems proves to be expensive, requiring elaborate parallelisation strategies.This paper presents two recent developments which aim at improving the parallel scalability of the quantum chemistry code ParaGauss. First, we introduce a new Fortran interface to parallel matrix algebra and its library implementation. This interface specifies a set of distributed data objects, combined with a set of linear algebra operators. Thus, complicated algebraic expressions can be expressed efficiently in pseudo-mathematical notation, while the numerical computations are carried out by back-end parallel routines. This technique is evaluated on relativistic transformations, as implemented in ParaGauss.The second development addresses the solution of the generalized matrix eigenvalue problem—an inherent step in electronic structure calculations. In the case the symmetry of a molecule is exploited, pertinent matrices expose a block-diagonal structure which makes the efficient use of existing parallel eigenvalue solvers difficult. We discuss a technique that uses a malleable parallel task scheduling (MPTS) algorithm for scheduling instances of parallel ScaLAPACK-routines on the available processor resources. This technique significantly improves the parallel performance of this numerical step, reducing the corresponding execution time to below 1 s in most applications considered.

KW - Density functional theory

KW - High performance computing

KW - Parallel numerical algebra

KW - Relativistic quantum chemistry

KW - Scheduling algorithms

UR - http://www.scopus.com/inward/record.url?scp=84899921315&partnerID=8YFLogxK

U2 - 10.1007/978-3-642-38762-3_6

DO - 10.1007/978-3-642-38762-3_6

M3 - Chapter

AN - SCOPUS:84899921315

T3 - Lecture Notes in Computational Science and Engineering

SP - 119

EP - 136

BT - Lecture Notes in Computational Science and Engineering

PB - Springer Verlag

ER -