Introduction

GAMESS-UK is an ab initio molecular electronic structure program for performing SCF- and MCSCF-gradient calculations, together with a variety of techniques for post Hartree Fock calculations. The program is currently available on a wide range of supercomputers, workstations and parallel machines.

The program provides for s,p,d,f and g-type Cartesian Gaussian orbitals, with open- and closed-shell SCF treatments available within both RHF and UHF frameworks. These treatments are augmented by generalised valence bond (GVB), complete active space SCF (CASSCF), and more general MCSCF calculations.  Ab initio core potentials are provided in both semi-local and non-local formalism for valence-only molecular orbital treatments.  The analytic energy gradient is available for each class of wavefunction above. Geometry optimisation may be performed in either internal or cartesian space, using a quasi-Newton rank-2 update method, while transition state location is available through either a synchronous transit, trust region or `hill-walking' method.  Force constants may be evaluated by either analytic methods (see below) or by numerical differentiation.  Configuration interaction estimates of the correlation energy may be generated through conventional-CI (using table-driven selection algorithms), Direct-CI and Full-CI treatments..

Coupled Hartree-Fock calculations of molecular polarisabilities and perturbations due to nuclear displacements allow for the analytic computation of all the dipole and quadrupole moment derivatives of a molecule.  Analytic second derivatives (force constants) of the energy, analytic calculations of polarisability derivatives and calculation of infrared and Raman intensities are also possible.  Moller-Plesset MP2, MP3 and MP4 perturbation theory calculations, include the analytic calculation of gradients, polarisabilities, dipole moment derivatives and force constants at the MP2 level.  Greens function OVGF and TDA methods are available for calculating ionisation spectra, while an RPA (Random Phase Approximation) module may be used in the study of molecular excitation spectra. Inter-fragment Energy Decomposition Analysis can be performed using the Morokuma approach.

A variety of features are introduced to augment present capabilities for treating excited states. These include both conventional and direct-RPA (random phase approximation) calculations of transition energies and oscillator strengths, and a MCLR (Multi-configurational Linear Response) module.

The treatment of correlation energy is enhanced through the inclusion of coupled cluster CCSD and CCSD(T) capabilities, due to Rendell and co-workers.

Considerable effort has been targeted to increasing both the range of computational methods available and the size of molecular system amenable to treatment.  The support and on-going development of the code has been greatly enhanced through the present arrangements between CCLRC and Computing for Science Ltd (CFS). Industrial funding has enabled CFS to appoint a support scientist at Daresbury, Jens Thomas, to focus on support issues and many of the developments, summarised below.  Jens’ recruitment has also enabled us to update the web site, and to start work on a revision of the build and QA processes, changes that should simplify distribution and maintenance in the future.

The code is presently licensed to some 226 academic and 7 industrial institutions world-wide. The current UK academic install base totals 84 sites.

Recent Developments

Many of the following features are discussed in more detail in a new paper describing the GAMESS-UK package [M. F. Guest, J. M.H. Thomas, P. Sherwood, I. J. Bush, H.J.J. van Dam, H. van Lenthe, R.W.A. Havenith and J. Kendrick, The GAMESS-UK Electronic Structure Package: Algorithms, Developments and Applications, Molec. Phys. (2005) in press].

• A major additional are of functionality is the provision of a Valence Bond Module through collaboration with the Utrecht group.
This is based on the Turtle code.
• Major enhancements to the DFT functionality, based on the CCP1 DFT module, including analytical second derivatives.
• A distributed-data FORTRAN95 module using MPI-based tools such as SCALAPACK has been added. This augments the parallel implementation based on the Global Array Tools, and was developed to overcome the poor scaling of the replicated data version of the code on the current high-end systems and to increase the size of calculations that can be undertaken by greater use of distributed data. All data structures, except those required for the Fock-matrix build are fully distributed. The implementation includes use of the recently developed "divide and conquer diagonaliser" from the ScaLAPACK library.
• Improved memory instrumentation has been added to ensure more efficient allocation and de-allocation of memory.
• Removal of 255 basis function limit in the MRDCI code. In addition Configuration Interaction calculations are now available for systems including g functions.
• The code has been ported to a range of new platforms, namely:  HP's (HPUX and f90) and IBM's (efc and Linux) Itanium-2 based servers and HP/Compaq's EV7 "marvel" servers. Enhancements to the 64-bit builds have focused on the platforms from Sun and IBM. Support has been added for both EM64T and Itanium2-based machines from Intel (Tiger Madison/1200 and 1500), HP (RX5670, RX4640 and RX2600), and SGI (the SGI Altix 3700), IBM's p-series turbo/1.7, AMDs 64-bitOpteron 244, 246, 840, 842 and 848, and Sun's V880/900-Cu, Macintosh OSX (G3, G4, and G5 processors) AMD Opteron and Athlon processors, Score parallel environment, Sun Server, Xeon processor with Myrinet interconnect, SGI Altix, Sunfire v880 server. A number of users (including the attendees at the DL codes workshop mentioned elsewhere) have emphasised the value of an efficient Windows port. We are now close to releasing a distribution of the version 6.3.2 code based on the IFC compiler - this work has included the deployment of the automatic installation tools now expected by users of this platform.  A free Macintosh demo version has been developed.
• TaskfarmingCapabilities: The parallel version of GAMESS-UK can now take advantage of the additional parallelism present in projects based on combinatorial methods in which a large number of smaller problems need to be solved at the same time. A taskfarming harness was developed for the Quantum Directed Virtual Evolution (QDVE) project that we are collaborating on with Marcus Durrant from the John Innes Centre in Norwich. The aim of QDVE is to develop an optimal catalyst for the conversion of nitrogen to hydrazine using an iterative genetic algorithm that characterises the best features of a random selection of catalysts and then creates a new batch of catalysts based on these features. The QDVE approach requires successive batches of roughly 500 small-molecule calculations (a "generation") to be run concurrently. The harness is an MPI program designed to be run on a large number of processors of a parallel machine. These processors are split into a number of equally sized groups, each of which is capable of running a single instance of the parallel version of GAMESS-UK. A single processor is set aside to act as a server process, reading in a list of the jobs to be processed and doling them out to the groups in a round-robin fashion. When a group has finished a job, it reports back to the server process, which hands it the next job to be run. When all of the jobs have completed and the groups all reported back, the server process brings down the job and prints a summary of the results.
• Developments to the Interface with CHARMM: An increasing number of requests for GAMESS-UK arise from workers in the biomolecular sciences who wish to use the code together with the CHARMM macromolecular modelling package.  The interface has seen a number of enhancements during the period of this report.  The interface with CHARMM has been optimised for use with DFT calculations, and some of the novel QM/MM coupling methods implemented in the code have been benchmarked in some detail. The most important new capability is the support for QM/MM pathway calculations using the Nudged Elastic Band (NEB) and Replica Path methodologies.
• Other Developments
• A new configure process is being developed, with the aim of simplifying the porting of the code to new platforms by localising all the machine-dependant parts of the process in a single, platform-dependant file.
• Optimisation of rotated axis integral code for basis sets which are not based on SP shell structure
• Changes to the CHPF solver to improve efficiency and reliability when used in the DFT case.
• Analysis of memory usage throughout the code with removal of bottlenecks.
• Checks on convergence of 2nd derivative accuracy with DFT grid size (ongoing)
• 2nd derivatives for systems employing ECP basis sets
• Thermodynamic analysis has been added in analytic Hessian runs.

Capabilities under Development

• Terascale algorithm development - a literature survey on approaches to parallel CCSD(T) has been undertaken and the implementation of a parallel module based on the work of Kobayashi and Rendell has started. This code is expected to form the basis for an exploration of different approaches to parallelisation of highly correlated methods (see section 11).
• Basis set optimisation (e.g. to optimise outer functions for negative ions)
• An option to consider atoms as frozen in geometry optimisation and Hessian analysis.
• Spin-orbit ZORA functionality, to support relativistic calculations including spin-orbit coupling.
• Preliminary planning for the implementation of a TD-DFT model has begun, as has discussion with possible collaborators on scientific applications of the module (W. Barford, Sheffield). This has included examination of the parallel implementation and memory use of the existing direct RPA program.
• The ADC code of Tarantelli and co-workers, (previously an IBM specific code) has been ported to the Intel platform in preparation for it's incorporation into GAMESS-UK.
• Further development of the Valence bond module has been carried out in collaboration with van Lenthe (Utrecht), features in development include

• improved orbital optimisation (improved pert) and stop criterion (still to be tested and to be part of thesis) This could make orbital optimisation on bigger molecules cheap, since the same number of matrix-elements are required as are needed in a Fock-matrix type optimisation.
• simple solvation models
• Availability of spherical harmonic basis sets

More Information

design by CCP1, December 2004