PNNL-13501 - Pacific Northwest National Laboratory

week on a 1 TFLOP/s computer with 1 gigabyte of
memory per processor. This leads to about
1,500 basis functions without symmetry and
3,000 with full symmetry. Larger calculations are
possible.
• Direct algorithm based upon the sequential semidirect
algorithm of Koch et al. (1994, 1996). Only
the amplitudes and iterative solution vectors are
stored on disk. The AO/SO integrals are recomputed
as necessary each CCSD iteration or for each pass of
the triples code. If sufficient memory/disk space is
available, the atomic orbital integrals can be cached.
• Many new capabilities were developed that can be
reused by other applications within NWChem
(integrals over symmetry adapted orbitals).
• The framework of the code is extensible to support
eventual integration of the Laplace triples and local
correlation methods.
• Detailed performance models support scaling of the
most expensive triples component to 10,000
processors with an expected efficiency of 99%.
Much effort has gone into achieving this since the
code does a large volume of communication. Some
of the techniques used are pre-sorting and blocking to
improve memory-locality and to bundle
communications, full dynamic load balancing, and
randomization to avoid communication hot spots.
The CCSD calculation will scale to a similar number
of processors, but not with such high efficiency
(perhaps 80% to 90%); however the CCSD portion
typically costs a factor of 5 to 10 less.
• Efficient execution on both massively parallel
computers and workstation clusters by avoiding
excess network traffic and reducing sensitivity to
latency.
• A new accelerated inexact Newton algorithm has
been formulated for solving the CCSD equations.
References
Corchado JC, Y-Y Chuang, PL Fast, J Villà, EL Coitiño,
W-P Hu, Y-P Liu, GC Lynch, KA Nguyen, CF Jackels,
MZ Gu, I Rossi, S Clayton, VS Melissas, R Steckler,
BC Garrett, AD Isaacson, and DG Truhlar. 1998.
POLYRATE-version 7.9.1. University of Minnesota,
Minneapolis.
110 FY 2000 Laboratory Directed Research and Development Annual Report
Kendall RA, E Aprà, DE Bernholdt, EJ Bylaska,
M Dupuis, GI Fann, RJ Harrison, J Ju, JA Nichols,
J Nieplocha, TP Straatsma, TL Windus, and AT Wong.
2000. “High performance computational chemistry; an
overview of NWChem a distributed parallel application.”
Computer Physics Communications 128, 260.
Koch H, O Christiansen, R Kobayashi, P Jorgensen, and
T Helgaker. 1994. “A direct atomic orbital driven
implementation of the coupled cluster singles and doubles
(CCSD) model.” CPL 228 233.
Koch H, A Sanchez de Meras, T Helgaker, and
O Christiansen. 1996. “The integral-direct coupled
cluster singles and doubles model.” JCP 104 4157.
Steckler R, W-P Hu, Y-P Liu, GC Lynch, BC Garrett,
AK Isaacson, D-h Lu, VS Melissas, TN Truong, SN Rai,
GC Hancock, JG Lauderdale, T Joseph, and DG Truhlar.
1995. POLYRATE-version 6.5. Computer Physics
Communications 88, 341-343.
Presentations
Nichols JA, DK Gracio, and J Nieplocha. 1999. “The
challenge of developing, supporting and maintaining
scientific simulation software for massively parallel
computers.” Germantown.
Seminar, “The Challenge of Developing, Supporting and
Maintaining Scientific Simulation Software for Massively
Parallel Computers.” San Diego Supercomputer Center,
San Diego, California, January 2000.
Plenary lecture, “NWChem: A MPP Computational
Chemistry Software Developed for Specific Application
Targets.” Spring 2000 Meeting of the Western States
Section of the Combustion Institute, Colorado School of
Mines, Golden, Colorado, March 2000.
FLC Awards Presentation, Charleston, South Carolina,
May 2000.
Invited lecture, “New Molecular Orbital Methods within
NWChem,” at the conference “Molecular Orbital Theory
for the New Millennium - Exploring New Dimensions
and Directions of the Molecular Orbital Theory.” Institute
of Molecular Science, Okazaki, Japan, January 2000.
“High Performance Computing in the Environmental
Molecular Science Laboratory.” Real World Computing
Institute, Tsukuba, Japan, January 2000.