TURBOMOLE

Turbomole is renowned for its computational highly efficient implementation
of Hartree-Fock (HF), Density Functional Theory (DFT), Coupled Cluster (CC)
and Moller-Plesset Perturbation Theory (MP). For excited states Time-Dependent DFT and response theory based CC methods are available.This is complemented by real-space DFT with periodic boundary conditions using the continuous fast multipole method (CFMM),two-component relativistic DFT, the cosmo solvens model, single-reference electron correlation methods (MP, CC) with explicit inter-electronic distance terms (r12/f12 methods) for more efficient treatment of electron correlation as well as the fast PNO (pair natural orbitals) variant of the CCSD method. Additionally, the availability of analytical and numerical first and second derivatives of the energy wrt geometric displacements and the support all point groups (in some methods restricted to Abelian point groups) and many more features make Turbomole an attractive package for the fast computation of arbitrary medium-sized molecules and clusters.

The initial target of the developers was to reliably deal with medium-sized to large molecules and clusters at moderate (or practically no) I/O demand using numerically stable methods of low computational complexity. As suchthe integral-direct approach is used throughout for the integrals in the atomic orbital basis as the most commonly used methods (HF, DFT, MP) can be formulated essentially in this basis. Together with today's standard tricks (auxiliary basis sets to reduce the formal scaling from N**4 to N**3, discarding integrals based on dynamic thresholds and many others) electronic structure calculations at this level are easy and straigtforward. Electron correlation methods on the other hand require many quantities in the molecular orbital basis. Four-index quantities are now mostly replaced by three-index quantities using auxiliary basis sets thereby reducing I/O and memory consumption by orders of magnitude.

The parallel scaling of Turbomole is case and method dependent. The old MPI /shared memory based ridft/rdgrad implementation developed ten years ago at JSC scales reasonably well given a sufficiently large problem since these modules are fast and invariably needed for structure optimizations. Since that time quite some effort has been spent on OpenMP parallelization of library modules so that nowadays most relevant modules are parallel. The amount of paralleli-zation and, hence, the observed speedup depends on the time spent in those parallelized library modules. This is case and method dependent and varies from modest 50% to well above 99%. While these numbers at first sight look impres-sive, the maximum parallel speedup is 2 to 100, respectively, provided an infinite number of processors is used, so that the time spend in the parallel code vanishes.

The Turbomole version provided by the JSC on the supercomputer systems for academic users is somewhat adapted to the JSC environment. Owing to the scalability issues and the large number of cores per node available it is frequently advisable to organize the work such, that multiple problems are executed simultaneously.

Official link: www.turbomol.org