SI2-SSI-1147794: Sustainable Development of Next-Generation Software in Quantum Chemistry
C. David Sherrill1, Edmond Chow2, and T. Daniel Crawford3
Acknowledgments
Serial PANACHE Application: DF-CCSD(T)
Interfaces to Quantum Chemistry Programs
OverviewWe seek to deploy rank-reduction approximations for the electron repulsion
integral tensor on massively-parallel hardware by developing an interface for the
abstract manipulation of factor tensors distributed using sustainable third-party
massively parallel APIs
Overall Toolbox: LibPANACHE
Density Fitting (DF)
Electron-Repulsion Integral (ERI) Tensor:
DF Factorization:
Three-Index DF Integrals:
Two-Index DF Integrals:
Symmetric DF Factorization:
1Center for Computational Molecular Science and Technology
School of Chemistry and Biochemistry
School of Computational Science and Engineering
Georgia Institute of Technology
2School of Computational Science and Engineering
Georgia Institute of Technology3Department of Chemistry
Virginia Polytechnic Institute and State University
Auxiliary Basis Expansion:
Interfaces Available:
• C++
• Fortran
Major Packages:
• PSI4
• GAMESS (in progress)
• Dalton (in progress)
Specific Programs:
• DFOCC: Orbital-optimized second-order many-body
perturbation theory with DF and CD (Ugur Bozkaya, Ataturk
University)
• FNOCC: DF/CD CCSD(T) with GPU capabilities (Eugene
DePrince, Florida State)
• DF-SAPT: DF/CD Symmetry-adapted perturbation theory with
natural orbital capabilities (Rob Parrish, Georgia Tech)
• CCENERGY: (DF-)CCSD(T) energies, gradients,
polarizabilities, excitation energies, chiroptical properties
(Daniel Crawford, Virginia Tech)
Conventional
Approximate
Serial Parallel
Integrals
Hardware
TargetAdvantages:
-Approximate integrals already provide
substantial tractability gains in serial
-Low-rank representations are easy to distribute
in aggregate memory
-Approximate integrals shift FLOP effort to
efficient, easily parallelized linear algebra
operations
Solver
CGSolver
DLSolver
DLXSolver
JK
PDirectJK
DFJK
DirectJK
V
RTDAV
RKSV
UKSV
Uses
Inherits
HF
TDHF
CPHF
CIS
SAPT0
Client Codes
LibFock
LibFock Layout:
Many-Body Problems: LibLowRank
LibLowRank Layout:
DF-ERI
CD-ERI
PS-ERI
Client Codes
TensorAPI
DF-SAPT0
PANACHE Layer
Serial/Parallel
InterfacesSerialTensor GlobalArrays CTF
Serial PANACHE Application: A-SAPT
Parallel PANACHE Application: GTFock
One-Body Problems: LibFock
2) Identify two key occupied bodies in
each many-body interaction term.
1) Allow the many-body interaction to
occur naturally (SAPT0 for now).
4) Visualize Results3) Perform key two-body summations
with local quasiparticles.
A-SAPT Formalism:
PANACHE Implementation:•>50 pages of complicated A-SAPT0 equations become 8200 lines of simple,
readable code (less than our previous SAPT0 code!)
•LibFock is used for electrostatics, exchange, induction
•LibLowRank is used for dispersion and some auxiliary computations
•Density fitting acceleration is automatically enabled by LibPANACHE
•Resultant test code is already production-level, works for >220 atoms
Application to Buckycatcher:Electrostatics:Exchange:Dispersion:
R.M. Parrish and C.D. Sherrill, J. Chem. Phys. 141, 044115; J. Chem. Theory Comput. 10, 4417 (2014)
A.E. DePrince III and C.D. Sherrill, J. Chem. Theory Comput., 9, 293 (2013)
A.E. DePrince III and C.D. Sherrill, J. Chem. Theory Comput., 9, 2687 (2013)
First Definitive Determination of 3-Body Contribution to the
Lattice Energy of an Organic Crystal
Errors of FNO-DF-CCSD(T) vs Conventional CCSD(T) for
Intermolecular Interactions (S22 Test Set)
DF/CD-FNO-CCSD(T) with T1-Transformation:
Represent all ERIs via DF/CD Approximation:
Work in a truncated natural orbital (FNO) virtual space:
Use a T1-Transformed Hamiltonian to reduce complexity; CCSD
equations then look like CCD
Examination of Indinivir Binding to HIV-II Protease
Using Truncated Models of Increasing Size
Parallel Integral-Direct Generalized J and K Matrices:
•2D tiling of shell pairs
•Predetermination of near-optimal task assignments
•Work-stealing construct for load-leveling
•Designed to minimize communications
•Built on GlobalArrays
PArallel Numerical Algorithms in CHemistry
Engine (PANACHE) Scheme:
Tensor
DF-CCSD(T)
Laplace-Den
THC-ERI
TensorOpsLibLowRank LibFock
JK
V
Solver
Uses
Inherits
Client Codes
LibLowRank
DF-ERI
CD-ERI
PS-ERI
Laplace-Den
THC-ERI
DF-MP2
DF-(T)
DF-ASAPT0
DF-CCSD
THC-CCSD
X
NO transformation
E. Chow, X. Liu, S. Misra, M. Dukhan, M. Smelyanskiy, J. R. Hammond, Y. Du, X.-K. Liao, and P. Dubey,
Int. J. High Performance Comput. Appl., submitted
E. Chow, X. Liu, M. Smelyanskiy, and J. R. Hammond, J. Chem. Phys., submitted
Demonstration of Negligible Errors of FNO (cutoff 10-5), Density Fitting, and Cholesky
Decomposition Approximations, and their combinations
•FNO-DF-CCSD(T) code allowed a resolution of the
controversy over the size of 3-body contributions to
the lattice energy of crystalline benzene [M. R.
Kennedy et al., J. Chem. Phys. 140, 121104 (2014)]
•3-body terms contribute 0.89 kcal/mol (7.2% of
lattice energy)
•3-body dispersion is only 0.76 kcal/mol,
significantly smaller than some recent estimates
•FNO-DF approach sped up trimer computations by
~4x (8.5 days to 2.1 days)
3.8 Å cutoff, 3555 bf
18.0 Å cutoff, 27394 bf,
879 TF/s Fock build on
Tianhe-2
GTFock vs NWChem on Stampede (3.8 Å system)