Parallel Computing—a route to complexity and reality in material simulations
Shiwu Gao
Department of Applied Physics
Chalmers/Göteborg University
Parallel computing and materials simulations
Water-metal interface
Dynamics of electron excitation/transfer
BiomembraneAquaporin water channel in membraneK. Murata et al, Nature, 407, 599 (2002)
Macroscopic(meter, hour)
Mesoscopic
Kinetics
Energetics
Atomic
Electronic(Å, fs)
Bottom-up approach
Theoretical approach based on:
1) Fundamental laws of physics
2) Computer modeling and simulations
Density Functional Theory based simulations
2
22
)()(
)()()()(2
Fi
i
iiiXCCoulomb
n
nvnm
rr
rr
Solving the Kohn-Sham Equations for all electorns
Full-potential and pseudopotential methods -Ful-potential methods (FP-LAPW, FP-LMTO) accurate and slow
-Pseudo-potential methods (VASP, CPMD, PWSCF) fast but with uncertainty in pseudopotentials
Outline • Parallelization of WIEN package
FP-(L)APW method
• Applications - Hydrogen bonding by CH group
- Pressure melting of confined water films
WIEN97 (T. U. Vienna)
Typical timing (s)
H/Cu(100) p(3x3) 3+2+5layers 29 atoms
Potential
Eigenproblem
Density
Core Electron
Mixing in/out data
+ Accurate + Versatile -- Slow-- larger RAM
Timing in LAPW1
0
10
20
30
40
50
60
70
80
90
H S Hns Solver
Exact
Iterative
- Large memory needed for H,S
RAM ~ M2
- Time-consuming
H |Ψk>= εkS |Ψk> t ~ M3
For large systems (>30 atoms)
- more than 90 % CPU time
- severval GB RAM
Parallelizing the eigenproblem (LAPW1)
2. Parallelizing the eigensolver
-Incorporating PQR
-Writting an iterative parallel solver
Myid = 0 1 2 3 0 1 2 3 0
1. Distributing and parallel setting H and S
PQR:X.B. Chi, Inst. Software, Chinese Academy of Sciences,Beijing
Further Parallelizations
+ LAPW1 Distributing H S setting and parallelizing the eigensolver -Incorporating PQR
-Writting an iterative parallel solver
+ LAPW2 and LAPW0
Distributing the calculation atom-wise
+ Implemeting the new APW+lo basis, E. Sjöstedt, Nordström, and Singh, Solid State Commun 114, 15 (2001)
S. Gao, Comput. Phys. Commun. (to be published)
Test example: C2H4+O2/Ag(110) coadsorption
- 100 surface atoms -Ag(110) 3x4x7=84
-(C2H4+o2)x2=16
- 6 layer vacuum- 21x23x35 au3
- Dual basis -Ag(110) LAPW -molecules, APW+lo
- 1-k point- 9 Ry cut-off structure- 13 -16 Ry in energy- 12 min/SCF 24 SGI3k- 12-15 Ionic steps/day
Scaling on Seth---Linux cluster at HPC2N
Up to 128 CPUs
Seth and SP3: 1) comparable scaling, 2) different in speed
Summary on scaling and performance
Timing consuming parts Acceleration on p CPUs
Setting H and S 0.98—1.0 p
HNS 0.79—0.9 p
Eigensolver--PQR 0.91–-0.94 p
Iterative Diag. 0.7— 0.8 p
Charge (LAPW2) ~ Na (or no acc.)
Potential (LAPW0) ~ Na (or no acc.)
Applicable to large systems, as PW-PP methods
Hydrogen bonding by CH group C2H4+O2/Ag(110)
Expt: J. R. Hahn, W. Ho, UCITheory: S. W. Gao, Chalmers
Why hydrogen bond with CH group
• H-bond is ubiquetous in biomolecules and organics
• Also of interest for fundamental studies (Ionic, covelency, vdW?)
• Usually with FH (VII), OH (VI), and NH(V) due to the large affinity, favoring ionic coupling
• H-bond with CH, weak—controversial EHB < 1 kcal/mol (c.a. 43 meV)
Distance-dependent interaction
-27.4 meV
-90.4 meV
-6.6 meV
In the gas phase: the interaction is negligible ~ + 10 meV
Background and Motivation
• Special phenomena in confined water• Bio-membrane fusion: role of thin water films• Pressure:
-phase control-material synthesis-mechanical stimuli in biology
• Ice-skating and lubrication• How to characterize confined liquid water from computer simulations
• New water phases in confined water
• Existence of solid-liquid critical points
K.Koga et al.,Nature 412, 802 (2001)
Simulation Method
• VASP—Veinna ab intio simulation package (better adapted to MD simulations)
.• Slab representation in a supercell
geometry: up to 48 Pt atoms and 32 H2O molecules
Summary
• Parallel WIEN for large-scale ab initio electron structure calculations
• Applications in material simulations1. Hydrogen bonding mechanism induced by
adsorption 2. Pressure induced phase transitions of water
films