An independent module for parallelization over a general class of tasks using a cooperative “work stealing” method was developed and implemented. Several variants were built by using various MPI-2 extensions. Among others, this module was validated for the two-electron integrals of the parallel density functional program ParaGauss. Together with a newly implemented convergence accelerator for the self-consistent field cycle this module resulted in a notable performance improvement of the program ParaGauss, especially for large systems and large numbers of processes. Furthermore, a library of global methods for transition state search was extended and improved, especially by a newly developed modified Dimer-Lanczos method. This library permits an efficient two-step strategy which was systematically tested for the first time on surface reactions.
«
An independent module for parallelization over a general class of tasks using a cooperative “work stealing” method was developed and implemented. Several variants were built by using various MPI-2 extensions. Among others, this module was validated for the two-electron integrals of the parallel density functional program ParaGauss. Together with a newly implemented convergence accelerator for the self-consistent field cycle this module resulted in a notable performance improvement of the program...
»