Документ взят из кэша поисковой машины. Адрес оригинального документа : http://classic.chem.msu.su/gran/gamess/largescale.pdf Дата изменения: Wed Nov 22 21:09:32 2006 Дата индексирования: Mon Oct 1 19:51:59 2012 Кодировка:

Large-scale QC modeling using PC GAMESS package.
Alex A. Granovsky

Laboratory of Chemical Cybernetics, M.V. Lomonosov Moscow State University, Moscow, Russia
June 30, 2006


Outline
Introduction to PC GAMESS Modeling of large systems:
High-level methods Linear scaling methods QM/MM-based approaches


The PC GAMESS project
Intel architecture specific high-performance parallel freely-available QC package Is being developed in the Laboratory of Chemical Cybernetics at MSU since 1993 Used by more than 2500 academic/industry research groups all around the world Supports Windows and Linux based SMP systems, clusters, and their combinations New state-of-the-art parallel algorithms for many QC calculation methods, scalable up to hundreds of nodes.


PC GAMESS vs. GAMESS US:
Strongly modified to achieve the maximum possible performance on Intel-based platforms; Functionally extended to provide QC methods which are not currently present in the regular GAMESS version; Designed to deal with large molecular systems efficiently; Runs much faster and requires less resources


Modeling of large systems, highlevel methods


High-level methods, current status
Routinely applicable to modeling of ground and excited states of medium-size systems:
State-of-the-art parallel algorithms for MP2MP4, MCSCF, and MCQDPT2 Allows to model medium size systems (up to 1000-4000 basis functions) with high accuracy and reasonable calculation time


High-level methods, examples


Medium size systems: Fullerene C60 and its dimer C120, MP2 calculations


Largest MP2 calculation attempted so far
System Basis Group N c n Nnodes Dynamic load balancing Real time data packing Asynchronous I/O Total FP operations count Distributed data size CPU time on master node, sec Wall clock time, sec. CPU usage, % Node performance, MFlops/s Performance, % of peak Cluster performance, GFlops/s C120 cc-pVTZ-f D2h 3000 120 240 18 on on off 3.321015 2.0 TB 89301 110826 80.5 1935 40.3 34.8

off off off 3.241015 2.0 TB 83029 150880 55 1330 27.7 23.9

on on on 3.3210 2.0 TB 95617 95130 100.5 2320 48.3 41.7

15

Pentium 4C 2.4 GHz / 1024MB / 120GB / Gigabit Ethernet


Medium size systems: structure and electronic spectra of retinal molecule in rhodopsin photoreceptor protein (visual pigment) 11-cis retinal chromophore
Protein Lys296 Protonated Shiff-base linkage

+

Rhodopsin ~ 5500 atoms
Glu113

12

11


Electronic spectra of retinal molecule in rhodopsin photoreceptor protein
Retinal molecule in protein (rhodopsin) environment - photosensitive receptor
77 atoms (H, C, N, O), 258 electrons approx. 5500 atoms in the protein

four-state QM/MM MCQDPT2 calculation for high-quality description of the excited states and electronic transition moments
Number of basis functions (N) 715 Number of FP operations ~ 1.5 · 1015

Less than 2 days on the workstation (this is one calculation of that kind

single dual-Nocona of the largest ever attempted so far).


High-level methods, future plans
No linear scaling approaches at present but will be implemented in the future:
Local correlation methods with linear scaling Resolution of Identity (RI)-based methods.


Modeling of large systems linear scaling methods for HF, DFT, CIS, TDHF, and TDDFT


Linear scaling methods, current status
Fast linear scaling Coulomb (J matrix) assembly via QFMM and near-field J engine State-of-the-art grid-based ExchangeCorrelation integration for linear scaling DFT and TDDFT State-of-the-art linear scaling HF exchange (K matrix) assembly for HF, hybrid DFT, CIS, TDHF, and TDDFT


Linear scaling methods, examples


Scaling example, Glycine oligomers (running on 18 nodes)
1000
GAMESS (US) PC GAMESS without Linear Scaling PC GAMESS Near J PC GAMESS Linear K

CPU Time, seconds

100

10

7

8

9

10

20

30
n

40

50

60

(Gly)


Fullerene (C60) 2D polymer (tetragonal phase) - modeled by PC GAMESS


Fullerene (C60) 2D polymer (hexagonal phase) - modeled by PC GAMESS


Linear scaling methods, future plans
Linear scaling QFMM energy gradients for HF, DFT, and TDDFT RI-based methods:
RI-J based code for pure DFT functionals RI-K based code for hybrid DFT Multipole Accelerated RI (MARI) energies and gradients for super-fast linear scaling DFT and TDDFT

GW-based methods for excited states and zone theory


Modeling of large systems QM/MM-based methods


QM/MM methods, current status
Several QM/MM approaches are already implemented in the PC GAMESS:
Mechanical embedding Electronic embedding

Different types of QM/MM:
Classical QM/MM (Jim Kress)
Several empirical force fields, including UFF

QM/DIM (Diatomics In Molecules) (A. Bochenkova)
Nonempirical many-body force fields

QM/EFP (GAMESS US)
Semiempirical and nonempirical corrections to H1


QM/MM methods, examples


Equilibrium structure of Retinal molecule in structure Rhodopsin protein (5500 atoms) protein


QM/MM methods - modeling effects of the environment


Equilibrium geometry of the retinal molecule in water solution - PC GAMESS QM/EFP results

Totally 275 water molecules


Equilibrium geometry of HArF molecule in argon matrix - PC GAMESS QM/DIM results

Totally 365 argon atoms


QM/MM methods, future plans

Implement polarizable force fields


The PC GAMESS on the Web:
http://classic.chem.msu.su/gran/gamess/index.html