Firefly and PC GAMESS/Firely LARGE-SCALE PARALLEL MCSCF CODE DOCUMENTATION Throughout this document we'll denote "large-scale" any CI, MCSCF, MCQDPT2, or XMCQDPT2 calculations of systems having large number of basis functions (e.g., 1500) and relatively small active spaces (e.g., 12 electrons/12 orbitals). In this case, the overall costs of calculations are mainly dominated by the integral transformation & effective Fock matrix construction steps. To speed up these stages, PC GAMESS/Firefly includes special fast direct and conventional integral transformation algorithms based on the fastints/gencon code. They have excellent parallel scalability and very modest memory requirements. They are available for FORS (CAS)-type CI (both GUGA & ALDET), SOSCF/FOCAS MCSCF, MCQDPT2, and XMCQDPT2 calculations only. Below is the summary of the most relevant options for high-performance large-scale CI/MCSCF/MCQDPT2/XMCQDPT2 jobs: 0. $contrl fstint=.t. gencon=.t. $end Use fastints/gencon code 1. $system kdiag=0 nojac=100 $end Instructs PC GAMESS/Firefly to use fast diagonalization routines if available and never use Jacobi diagonalization for matrices of size 100x100 and above. 2. $p2p p2p=.t. dlb=.t. $end Instructs PC GAMESS/Firefly to use dynamic load balancing over p2p interface during parallel runs. This is the best strategy, although static load balancing will work as well. 3. $trans mptran=2 dirtrf=.t. aoints=dist altpar=.t. mode=gsm $end Instructs PC GAMESS/Firefly to select fast alternative algorithm for integral transformation, using either its direct variant (preferred), or conventional with 2-e integrals distributed over nodes, and selects alternative (more scalable) parallel strategy for MSCSF runs. gsm is 3-digit decimal number defining the details of the direct parallel transformation code to be used. g can be either 0 or 1, and means either to use (1) or not to use (0) gencon version of the fastints code. s can be either 0 or 1, and means either to use (1) or not to use (0) approximate Schwarz integral screening (note that even if approximate screening is disabled, the exact Schwarz screening will be nevertheless in effect by default). Finally, m can be 0, 1, and 2, and means small (0), medium (1), or large (2) active space. Thus, mode=112 is the most appropriate for most runs, but for very small active spaces, mode=110 or 111 will perform faster. 4. $ciinp castrf=.t. $end Selects fast MCSCF-like integral transformation for standalone CI runs. $transt castrf=.t. $end Selects fast MCSCF-like integral transformation for standalone CI transtion moment/spinorbit runs. 5. $gugem pack2=.t. $end Selects additional packing of GUGA CI matrix for GUGA-style CI or MCSCF. 6. $mcscf fullnr=.f. soscf=.t. $end or $mcscf fullnr=.f. soscf=.f. focas=.t. $end Selects supported fast MCSCF algorithms (note fullnr does not presently support fastints/gencon).
Last updated: February 7, 2010