Performance and Productivity: NWChem

1

Performance and Productivity: NWChem

Robert J. HarrisonOak Ridge National Laboratory,

University of Tennessee

NWChem is funded by the U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research, under contract DE-AC06-76RLO 1830 with Battelle Memorial Institute (Pacific Northwest National Laboratory, PNNL) as part of the Environmental Molecular Sciences Laboratory, PNNL.

2

R. J. Harrison, J. A. Nichols, T. P. Straatsma, M. Dupuis,E. J. Bylaska, G. I. Fann, T. L. Windus, E. Apra, W. de Jong,S. Hirata, M. T. Hackler, J. Anchell, D. Bernholdt, P. Borowski,T. Clark, D. Clerc, H. Dachsel, M. Deegan, K. Dyall, D. Elwood,H. Fruchtl, E. Glendening, M. Gutowski, K. Hirao, A. Hess,J. Jaffe, B. Johnson, J. Ju, R. Kendall, R. Kobayashi, R. Kutteh,Z. Lin, R. Littlefield, X. Long, B. Meng, T. Nakajima,J. Nieplocha, S. Niu, M. Rosing, G. Sandrone, M. Stave, H. Taylor,G. Thomas, J. van Lenthe, K. Wolinski, A. Wong, and Z. Zhang,

"NWChem, A Computational Chemistry Package for Parallel

Computers, Version 4.1" (2002),

Pacific Northwest National Laboratory,Richland, Washington 99352-0999, USA.

NWChem CitationT.H. Dunning, Jr., D.A. Dixon, M.F. Guest

3

NWChem Overview• Provides major new modeling and simulation capability for

molecular science– Broad range of molecules, including biomolecules

– Electronic structure of molecules (non-relativistic, relativistic, one-/two-component, ECPs, second deriv.)

– Increasingly extensive solid state capability

– Molecular dynamics, molecular mechanics

• Extensible and long-lived

• Freely distributed – installed at about 1000 sites worldwide

• Performance characteristics – designed for MPP– Single node performance comparable to best serial codes

– Scalability to 1000’s of processors

• Portable – runs on a wide range of computers

4

Molecular Science Software Suite (MS3)

http://www.emsl.pnl.gov/pub/docs/ecce/

http://www.emsl.pnl.gov/pub/docs/parsoft/http://www.emsl.pnl.gov/pub/docs/nwchem/

5

6

7

NWChem to go ...Compaq iPAQ

Linux (Intimate)

64 Mbyte RAM,

16 Mbyte flash,

2 Gbyte

PCMCIA disk

Strongarm CPU

Sadly, no FPU

8

Higher-level composition

• Modular, hierarchical design – Easy to access high level features– Easy to extend with new high level features

• Standardized interfaces– Reuse of low-level functionality without side effects

• Distributed-shared memory parallel programming model– Non-uniform memory access (NUMA) aware algorithms

• Python interface– User’s & developers can write NWChem programs in Python

• Cool stuff – now becoming available– Automatic code generation – compose with many-body theory and/or

tensor expressions (already in NWChem 4.5)– Multiresolution quantum chemistry – compose with operators and

functions

9

NWChem Architecture

Ru

n-t

ime d

ata

base

DFT energy, gradient, …

MD, NMR, Solvation, …

Optimize, Dynamics, …

SCF energy, gradient, …

Inte

gra

l A

PI

Geom

etr

y

Ob

ject

Basis

Set

Ob

ject

...

PeIG

S

...

Global Arrays

Memory Allocator

Parallel IO

MolecularModelingToolkit

MolecularCalculation

Modules

MolecularSoftware

DevelopmentToolkit

GenericTasksEnergy, structure, …

• Object-oriented design– abstraction, data hiding, APIs

• Parallel programming model– non-uniform memory access, global

arrays, MPI

• Infrastructure– GA, Parallel I/O, RTDB, MA, ...

• Program modules– communication only through the

database– persistence for easy restart

10

Issues in Parallel Computing

• Expressing and managing concurrency

• The memory hierarchy

• Efficient sequential execution

11

The Memory Hierarchy• Non-uniform memory access - NUMA

– Your workstation is NUMA - registers, cache, main memory, virtual memory

– Parallel computers just add non-local memory(s)– Unites sequential and parallel computation

• Differ only in expression and management of concurrency

• Distributed data– Do not limit calculation by resources of one node– Exploit aggregate resources of the whole machine– SCF and DFT can distribute all data > O(N)– MP2 gradients distribute all data > O(N2)

12

Parallel Programming ModelGlobal Arrays + MPI

• J. Nieplocha• Supported by DOE/ASCR/MICS• Shared-memory-like model

– Fast local access– NUMA aware and easy to use– MIMD and data-parallel modes– Inter-operates with MPI, …

• BLAS and linear algebra interface• Used by most major chemistry codes,

also in financial futures forecasting, astrophysics, computer graphics, …

• Ported to major parallel machines– IBM, Cray, SGI, clusters, ...

Single, shared data structure

Physically distributed data

http://www.emsl.pnl.gov/pub/docs/global

13

local memory

Non-uniform memory access model of computation

Shared Object

copy to local mem

ory

Shared Object

cop

y to

sha

red

obje

ct

local memorylocal memory

compute/update

1-sidedcommunication

1-sidedcommunication

14

O(1) programmers … O(1000) nodes …

O(100,000) processors … O(10,000,000) threads

• Expressing/managing concurrency at the petascale– It is too trite to say that the parallelism is in the physics

– Must express and discover parallelism at more levels

– Low level tools (MPI, Co-Array Fortran, UPC, …) don’t discover parallelism or hide complexity or facilitate abstraction

• Management of the memory hierarchy– Sending data from one multiprocessor chip to another will be

like us taking a trip to Europe

– Memory will be deeper ; less uniformity between vendors

– Need tools to automate and manage this, even at runtime

15

Synthesis of High Performance Algorithms for Electronic Structure

Calculations• Sadayappan, Baumgartner, Cociorva, Pitzer (OSU)

Ramanujam (LSU)Bernholdt, Dean, White III, Harrison (ORNL)Hirata (PNNL)Nooijen (Waterloo)

• Objective:

– Automate the implementation of optimized parallel computer programs for many-electron methods expressed as tensor contractions

• Multi-disciplinary, multi-institution project

• Collaboration between NSF ITR, DOE SciDAC, and ORNL LDRD projects

16

CCSD Doubles Equationhbar[a,b,i,j] == sum[f[b,c]*t[i,j,a,c],{c}] -sum[f[k,c]*t[k,b]*t[i,j,a,c],{k,c}] +sum[f[a,c]*t[i,j,c,b],{c}] -sum[f[k,c]*t[k,a]*t[i,j,c,b],{k,c}]

-sum[f[k,j]*t[i,k,a,b],{k}] -sum[f[k,c]*t[j,c]*t[i,k,a,b],{k,c}] -sum[f[k,i]*t[j,k,b,a],{k}] -sum[f[k,c]*t[i,c]*t[j,k,b,a],{k,c}] +sum[t[i,c]*t[j,d]*v[a,b,c,d],{c,d}] +sum[t[i,j,c,d]*v[a,b,c,d],{c,d}] +sum[t[j,c]*v[a,b,i,c],{c}] -sum[t[k,b]*v[a,k,i,j],{k}] +sum[t[i,c]*v[b,a,j,c],{c}] -sum[t[k,a]*v[b,k,j,i],{k}] -sum[t[k,d]*t[i,j,c,b]*v[k,a,c,d],{k,c,d}] -sum[t[i,c]*t[j,k,b,d]*v[k,a,c,d],{k,c,d}] -sum[t[j,c]*t[k,b]*v[k,a,c,i],{k,c}] +2*sum[t[j,k,b,c]*v[k,a,c,i],{k,c}] -sum[t[j,k,c,b]*v[k,a,c,i],{k,c}] -sum[t[i,c]*t[j,d]*t[k,b]*v[k,a,d,c],{k,c,d}] +2*sum[t[k,d]*t[i,j,c,b]*v[k,a,d,c],{k,c,d}] -sum[t[k,b]*t[i,j,c,d]*v[k,a,d,c],{k,c,d}] -sum[t[j,d]*t[i,k,c,b]*v[k,a,d,c],{k,c,d}] +2*sum[t[i,c]*t[j,k,b,d]*v[k,a,d,c],{k,c,d}] -sum[t[i,c]*t[j,k,d,b]*v[k,a,d,c],{k,c,d}] -sum[t[j,k,b,c]*v[k,a,i,c],{k,c}] -sum[t[i,c]*t[k,b]*v[k,a,j,c],{k,c}] -sum[t[i,k,c,b]*v[k,a,j,c],{k,c}] -sum[t[i,c]*t[j,d]*t[k,a]*v[k,b,c,d],{k,c,d}] -sum[t[k,d]*t[i,j,a,c]*v[k,b,c,d],{k,c,d}] -sum[t[k,a]*t[i,j,c,d]*v[k,b,c,d],{k,c,d}] +2*sum[t[j,d]*t[i,k,a,c]*v[k,b,c,d],{k,c,d}] -sum[t[j,d]*t[i,k,c,a]*v[k,b,c,d],{k,c,d}] -sum[t[i,c]*t[j,k,d,a]*v[k,b,c,d],{k,c,d}] -sum[t[i,c]*t[k,a]*v[k,b,c,j],{k,c}] +2*sum[t[i,k,a,c]*v[k,b,c,j],{k,c}] -sum[t[i,k,c,a]*v[k,b,c,j],{k,c}] +2*sum[t[k,d]*t[i,j,a,c]*v[k,b,d,c],{k,c,d}] -sum[t[j,d]*t[i,k,a,c]*v[k,b,d,c],{k,c,d}] -sum[t[j,c]*t[k,a]*v[k,b,i,c],{k,c}] -sum[t[j,k,c,a]*v[k,b,i,c],{k,c}] -sum[t[i,k,a,c]*v[k,b,j,c],{k,c}] +sum[t[i,c]*t[j,d]*t[k,a]*t[l,b]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[k,b]*t[l,d]*t[i,j,a,c]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[k,a]*t[l,d]*t[i,j,c,b]*v[k,l,c,d],{k,l,c,d}] +sum[t[k,a]*t[l,b]*t[i,j,c,d]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[j,c]*t[l,d]*t[i,k,a,b]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[j,d]*t[l,b]*t[i,k,a,c]*v[k,l,c,d],{k,l,c,d}] +sum[t[j,d]*t[l,b]*t[i,k,c,a]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[i,c]*t[l,d]*t[j,k,b,a]*v[k,l,c,d],{k,l,c,d}] +sum[t[i,c]*t[l,a]*t[j,k,b,d]*v[k,l,c,d],{k,l,c,d}] +sum[t[i,c]*t[l,b]*t[j,k,d,a]*v[k,l,c,d],{k,l,c,d}] +sum[t[i,k,c,d]*t[j,l,b,a]*v[k,l,c,d],{k,l,c,d}] +4*sum[t[i,k,a,c]*t[j,l,b,d]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[i,k,c,a]*t[j,l,b,d]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[i,k,a,b]*t[j,l,c,d]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[i,k,a,c]*t[j,l,d,b]*v[k,l,c,d],{k,l,c,d}] +sum[t[i,k,c,a]*t[j,l,d,b]*v[k,l,c,d],{k,l,c,d}] +sum[t[i,c]*t[j,d]*t[k,l,a,b]*v[k,l,c,d],{k,l,c,d}] +sum[t[i,j,c,d]*t[k,l,a,b]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[i,j,c,b]*t[k,l,a,d]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[i,j,a,c]*t[k,l,b,d]*v[k,l,c,d],{k,l,c,d}] +sum[t[j,c]*t[k,b]*t[l,a]*v[k,l,c,i],{k,l,c}] +sum[t[l,c]*t[j,k,b,a]*v[k,l,c,i],{k,l,c}] -2*sum[t[l,a]*t[j,k,b,c]*v[k,l,c,i],{k,l,c}] +sum[t[l,a]*t[j,k,c,b]*v[k,l,c,i],{k,l,c}] -2*sum[t[k,c]*t[j,l,b,a]*v[k,l,c,i],{k,l,c}] +sum[t[k,a]*t[j,l,b,c]*v[k,l,c,i],{k,l,c}] +sum[t[k,b]*t[j,l,c,a]*v[k,l,c,i],{k,l,c}] +sum[t[j,c]*t[l,k,a,b]*v[k,l,c,i],{k,l,c}] +sum[t[i,c]*t[k,a]*t[l,b]*v[k,l,c,j],{k,l,c}] +sum[t[l,c]*t[i,k,a,b]*v[k,l,c,j],{k,l,c}] -2*sum[t[l,b]*t[i,k,a,c]*v[k,l,c,j],{k,l,c}] +sum[t[l,b]*t[i,k,c,a]*v[k,l,c,j],{k,l,c}] +sum[t[i,c]*t[k,l,a,b]*v[k,l,c,j],{k,l,c}] +sum[t[j,c]*t[l,d]*t[i,k,a,b]*v[k,l,d,c],{k,l,c,d}] +sum[t[j,d]*t[l,b]*t[i,k,a,c]*v[k,l,d,c],{k,l,c,d}] +sum[t[j,d]*t[l,a]*t[i,k,c,b]*v[k,l,d,c],{k,l,c,d}] -2*sum[t[i,k,c,d]*t[j,l,b,a]*v[k,l,d,c],{k,l,c,d}] -2*sum[t[i,k,a,c]*t[j,l,b,d]*v[k,l,d,c],{k,l,c,d}] +sum[t[i,k,c,a]*t[j,l,b,d]*v[k,l,d,c],{k,l,c,d}] +sum[t[i,k,a,b]*t[j,l,c,d]*v[k,l,d,c],{k,l,c,d}] +sum[t[i,k,c,b]*t[j,l,d,a]*v[k,l,d,c],{k,l,c,d}] +sum[t[i,k,a,c]*t[j,l,d,b]*v[k,l,d,c],{k,l,c,d}] +sum[t[k,a]*t[l,b]*v[k,l,i,j],{k,l}] +sum[t[k,l,a,b]*v[k,l,i,j],{k,l}] +sum[t[k,b]*t[l,d]*t[i,j,a,c]*v[l,k,c,d],{k,l,c,d}] +sum[t[k,a]*t[l,d]*t[i,j,c,b]*v[l,k,c,d],{k,l,c,d}] +sum[t[i,c]*t[l,d]*t[j,k,b,a]*v[l,k,c,d],{k,l,c,d}] -2*sum[t[i,c]*t[l,a]*t[j,k,b,d]*v[l,k,c,d],{k,l,c,d}] +sum[t[i,c]*t[l,a]*t[j,k,d,b]*v[l,k,c,d],{k,l,c,d}] +sum[t[i,j,c,b]*t[k,l,a,d]*v[l,k,c,d],{k,l,c,d}] +sum[t[i,j,a,c]*t[k,l,b,d]*v[l,k,c,d],{k,l,c,d}] -2*sum[t[l,c]*t[i,k,a,b]*v[l,k,c,j],{k,l,c}] +sum[t[l,b]*t[i,k,a,c]*v[l,k,c,j],{k,l,c}] +sum[t[l,a]*t[i,k,c,b]*v[l,k,c,j],{k,l,c}] +v[a,b,i,j]

17

TCE Components

• Algebraic Transformations– Minimize operation count

• Memory Minimization– Reduce intermediate storage

• Space-Time Transformation– Trade storage for recomputation

• Storage Management and Data Locality Optimization– Optimize use of storage

hierarchy

• Data Distribution and Partitioning– Optimize parallel layout

Tensor Expressions

Algebraic Transformations

Memory Minimization

Performance Model

System Memory

Specification

Software Developer

Data Distribution and Partitioning

Parallel CodeFortran/C/…

OpenMP/MPI/Global Arrays

Sequence of Matrix ProductsElement-wise Matrix Operations

Element-wise Function Eval.

Space-Time Trade-Offs

Storage and Data Locality Management

No sol’n fits disk Sol’n fits disk, not mem. Sol’n fits mem.

Sol’n fits mem.

No sol’n fits disk

18

Multiresolution Quantum Chemistry

Robert J. Harrison, George I. Fann, Takeshi Yanai, Zhengting Gan

Oak Ridge National Laboratory andUniversity of Tennessee, Knoxville

andGregory BeylkinUniversity of Colorado

harrisonrj@ornl.gov

19

The funding

• This work is funded by the U.S. Department of Energy, the division of Basic Energy Science, Office of Science, under contract DE-AC05-00OR22725 with Oak Ridge National Laboratory. This research was performed in part using – the Molecular Science Computing Facility in the Environmental Molecular

Sciences Laboratory at the Pacific Northwest National Laboratory under contract DE-AC06-76RLO 1830 with Battelle Memorial Institute,

– resources of the National Energy Scientific Computing Center which is supported by the Office of Energy Research of the U.S. Department of Energy under contract DE-AC03-76SF0098,

– and the Center for Computational Sciences at Oak Ridge National Laboratory under contract DE-AC05-00OR22725 .

• ORNL LDRD

20

Outline• Brief introduction to methodology• Practical computation in higher dimensions

– Separated form for operators

• Analytic derivatives– Initial results

• Accuracy, timing and scaling• MP2

– Path to basis set limit results?

21

Objectives• Complete elimination of the basis error

– One-electron models (e.g., HF, DFT)– Pair models (e.g., MP2, CCSD, …)

• Correct scaling of cost with system size• General approach

– Readily accessible by students and researchers– Much smaller computer code than Gaussians– No two-electron integrals – replaced by fast application of

integral operators

• Fast algorithms with guaranteed precision

22

References

• The (multi)wavelet methods in this work are primarily based upon– Alpert, Beylkin, Grimes, Vozovoi (J. Comp. Phys., in press)– B. Alpert (SIAM Journal on Mathematical Analysis 24, 246-262, 1993). – Beylkin, Coifman, Rokhlin (Communications on Pure and Applied

Mathematics, 44, 141-183, 1991.)

• The following are useful further reading– Daubechies, “Ten lectures on wavelets”– Walnut, “An introduction to wavelets”– Meyer, “Wavelets, algorithms and applications”– Burrus et al, “Wavelets and Wavelet transforms”

23

Linear Combination of Atomic Orbitals (LCAO)

• Molecules are composed of (weakly) perturbed atoms– Use finite set of atomic wave functions as the basis– Hydrogen-like wave functions are exponentials

• E.g., hydrogen molecule (H2)

• Smooth function ofmolecular geometry

• MOs: cusp at nucleuswith exponential decay

1 ( ) e

( ) e e

r

r a r b

s r

r

0

0.2

0.4

0.6

0.8

1

1.2

1.4

-2 -1 0 1 2

24

LCAO• A fantastic success, but …• Basis functions have extended support

– causes great inefficiency in high accuracy calculations

– origin of non-physical density matrix

• Basis set superposition error (BSSE)– incomplete basis on each center leads to over-binding as

atoms are brought together

• Linear dependence problems– accurate calculations require balanced approach to a

complete basis on every atom

• Must extrapolate to complete basis limit– unsatisfactory and not feasible for large systems

25

Why “think” multiresolution?• It is everywhere in nature/chemistry/physics

– Core/valence; high/low frequency; short/long range; smooth/non-smooth; atomic/nano/micro/macro scale

• Common to separate just two scales– E.g., core orbital heavily contracted, valence flexible– More efficient, compact, and numerically stable

• Multiresolution– Recursively separate all length/time scales– Computationally efficient and numerically stable– Coarse-scale models that capture fine-scale detail

26

How to “think” multiresolution• Consider a ladder of function spaces

– E.g., increasing quality atomic basis sets, or finer resolution grids, …

• Telescoping series

– Instead of using the most accurate representation, use the difference between successive approximations

– Representation on V0 small/dense; differences sparse

– Computationally efficient; possible insights

0 1 2 nV V V V

0 1 0 2 1 1( ) ( ) ( )n n nV V V V V V V V

27

Scaling Function Basis• Divide domain into 2n pieces (level n)

– Adaptive sub-division (local refinement)– lth sub-interval [l*2-n,(l+1)*2-n] l=0,…,n-1

• In each sub-interval define a polynomial basis– First k Legendre polynomials– Orthonormal, disjoint support

n=0

n=1

n=2

l=0 l=1 l=2 l=3

/ 2

( ) 2 1 (2 1)

( ) 2 (2 )

i i

n n nil i

x i P x

x x l

28

Scaling Function Basis - II21( ) , 0, ,3i x i

i=0

i=2i=3

i=1

29

Multiwavelet Basis

• Space of polynomials on level n is Vn

• Wavelets - an orthonormal basis to span • Currently use Alpert’s basis• Vanishing moments

– Critically important property

– Since Wn is orthogonal to Vn the first k moments of functions in Wn vanish, i.e.,

• Sparse representations of many physically important kernels

1n n nW V V

( ) 0, 0, , 1jix x dx j k

30

Some Consequences of Vanishing Moments

• Compact representation of smooth functions– Consider Taylor series … the first k terms vanish and

smooth implies higher order terms are small

• Compact representation of integral operators– E.g., 1/|r-s| – Consider double Taylor series or multipole expansion– Interaction between wavelets decays as r-2k-1

• Derivatives at origin vanish in Fourier space– Diminishes effect of singularities at that point

31

• Slice thru grid used to represent the nuclear potential for H2 using k=7 to a precision of 10-5.

• Automatically adapts – it does not know a priori where the nuclei are.

• Nuclei at dyadic points on level 5 – refinement stops at level 8

• If were at non-dyadic points refinement continues (to level ??) but the precision is still guaranteed.

• In future will unevenly subdivide boxes to force nuclei to dyadic points.

32

33

Integral Formulation

• E.g., used by Kalos, 1962

212

12

2

2 2

2 *

* ( ) ( ) in 3D ; 24

k r s

V E

E V

G V

eG f r ds f s k E

r s

34

Integral operators in 3D

• Non-standard matrix elements easy to evaluate from compressed form of kernel K(x)

• Application in 1-d is fairly efficient– O(Nboxk2) operations

• In 3-d seems to need O(Nboxk6) operations– Prohibitively expensive

• Separated form– Beylkin, Cramer, Mohlenkamp, Monzon– O(Nboxk4) or better in 3D

* ( ) ( )T f dsK r s f s

', ', 'n l

ii jj kkr

'nl

iir

35

Low Separation Rank Representation

• Many functions/operators have short expansions

• Different from low operator rank– E.g., identity has full operator rank, but unit separation rank.

( )1

1 1

( )

2

( , , ) ( ) ( )

1

0

M dl

d l i il i

li

l

f x x f x O

f

36

Separated form for integral operators

• Approach in current prototype code– Represent the kernel over a finite range as a sum of Gaussians

– Only need compute 1D transition matrices (X,Y,Z)

– SVD the 1-D operators (low rank away from singularity)

– Apply most efficient choice of low/full rank 1-D operator

– Even better algorithms not yet implemented

2

', ', ' ' ' '

( ) ( )

yx z

i

nlnl nlnlii jj kk i ii ii ii

i

t ri

i

r X Y Z

K r e O

* ( ) ( )T f dsK r s f s

37

Accurate Quadratures

• Trapezoidal quadrature– Geometric precision for

periodic functions with sufficient smoothness.

2 2 2 2

2 2 2 2

/ 4

0

/ 4

2

2 s s

rx t t

x e e s

ee dt

r

e ds

The kernel for x=1e-4,1e-3,1e-2,1e-,1e0. The curve for x=1e-4 is the rightmost

38

Automatically generated representations ofexp(-30r)/r accurate to 1e-10, 1e-8, 1e-6, 1e-4, and 1e-2 (measured by the weighted error r(exp(-30r)/r - fit(r))) for r in [1e-8,1] were formed with 92, 74, 57, 39 and 21 terms, respectively.

Note logarithmic dependence upon precision.

39

Smoothed Nuclear Potential

2 24

2

0

erf 1( ) 16

3

1( ) 0 for 0,1,2

r r

n

ru r e e

r

drr u r nr

• u(r/c)/c shifts error to r<c• =0.00435*Z5*c3

• <V> accurate• <T> main source of error

40

Translational Invariance

• Dyadic

10-3 -75.9139

10-5 -75.913564

10-7 -75.91355634

• Non-dyadic

-75.9139

-75.913564

-75.91355635

• Uncontracted aug-cc-pVQZ –75.913002• Solving with e=1e-3, 1e-5, 1e-7 (k=7,9,11)• Demonstrates translation invariance and that forcing to dyadic

points is only an optimization and does not change the obtained precision.

• Average orbital sizes 1.6Mb, 8Mb, 56Mb

41

Analytic Derivatives

• Hellman-Feynman theorem applies

E VO

q q

42

N2 Hartree-Fock R=2.0 a.u.Basis Grad.Err. EnergyErr.

cc-pVDZ 5e-2 4e-2

aug-cc-pVDZ 5e-2 4e-2

cc-pVTZ 7e-3 1e-2

aug-cc-pVTZ 6e-3 9e-3

cc-pVQZ 8e-4 2e-3

aug-cc-pVQZ 9e-4 2e-3

cc-pV5Z 1e-4 4e-4

aug-cc-pV5Z 2e-5 2e-4

k=5 6e-3 1e-2

k=7 4e-5 2e-5

k=9 3e-7 -2e-7

k=11 0.0 0.0

0.026839623 -108.9964232

43

Sources of error in the gradient• Partially converged orbitals

– Same as for “conventional” methods

• Smoothed potential• Numerical errors in the density/potential

– Higher-order convergence except where the functions are not sufficiently smooth

• Inadequate refinement (clearly adequate for the energy, but not necessarily for other properties)

• Exacerbated by nuclei at non-dyadic points• Gradient measures loss of spherical symmetry around the nucleus … the

large value of the derivative potential amplifies small errors

44

Dependence onpotential smoothingparameter (c)

Absolute errors ofderivatives for diatomics with thenuclei at dyadic points.

For energy accuracyof 1e-6H 0.039Li 0.0062B 0.0026N 0.0015O 0.0012F 0.00099

45

Dependence onpotential smoothingparameter (c)

Absolute errors ofderivatives for diatomics with thenuclei at non-dyadic points.

For energy accuracyof 1e-6H 0.039Li 0.0062B 0.0026N 0.0015O 0.0012F 0.00099

46

Comparison with NUMOL and aug-cc-pVTZ

• H2, Li2, LiH, CO, N2, Be2, HF, BH, F2, P2, BH3, CH2, CH4, C2H2, C2H4, C2H6, NH3, H2O, CO2, H2CO, SiH4, SiO, PH3, HCP

• NUMOL, Dickson & Becke JCP 99 (1993) 3898

• Dyadic points (0.001a.u.) + Newton correction

• Agrees with NUMOL to available precision– LDA (k=7,0.002; k=9, 0.0006)

• k=9 vs. aug-cc-pVTZ rms error– Hartree-Fock 0.004 a.u. (0.019 SiO)– LDA 0.003 a.u. (0.018 SiO)

47

High-precision Hartree-Fock geometry for water

• Pahl and Handy Mol. Phys. 100 (2002) 3199– Plane waves + polynomials for the core

– Finite box (L=18) requires extrapolation

– Estimated error 3H, 1e-5 Angstrom

• k=11, conv.tol=1e-8,=1e-9, L=40– Max. gradient = 3e-8, RMS step=5e-8

– Difference to Pahl 10H, 4e-6 Angstrom, 0.0012

Basis OH HOH Energy

k=11 0.939594 106.3375 -76.06818006

Pahl 0.939598 106.3387 -76.068170

cc-pVQZ 0.93980 106.329 -76.066676

48

Energy Timing

• Water LDA with energy error of 1e-5– Initial prototype code with lots of Python overhead

• 450s on 2.4 GHz Pentium IV processor

– Current version (revised tensor class, integral operators)

• 96s on 2.4 GHz Pentium IV processor

• Predicted future performance– < 30s with known algorithmic improvements– faster still with better representations of the separated

operators, alternative basis sets, improved iterative solution

49

Asymptotic Scaling• Current implementation

– Based upon canonical orbitals – O(N) to O(N2) currently dominant (+ O(N3) linear algebra)

• Density matrix/spectral projector– Well established – O(Natomlogm()) to any finite precision

(Goedecker, Beylkin, …)– This is not possible with conventional AO Gaussians– Need separated representation for efficiency

• Gradient– each dV/dx requires O(-log()log(vol.)) terms

– All gradients evaluated in O(-Natomlog()log(vol.))

50

Water dimer LDAaug-cc-pVTZ geometry, kcal/mol.

Basis Uncorrected BSSE Corrected

cc-pVDZ -11.733 -3.958 -7.775

cc-pVTZ -9.464 -1.654 -7.810

cc-pVQZ -8.708 -0.821 -7.888

aug-cc-pVDZ -8.187 -0.382 -7.805

aug-cc-pVTZ -7.992 -0.086 -7.906

aug-cc-pVQZ -7.995 -0.054 -7.941

=10-3 -6.483

=10-5 -7.932

=10-7 -7.943

51

Benzene dimer LDAaug-cc-pVDZ geometry, kcal/mol.

Basis Uncorrected BSSE Corrected

cc-pVDZ -1.506 -1.035 -0.471

cc-pVTZ -1.271 -0.387 -0.884

cc-pVQZ -1.074 -0.193 -0.881

aug-cc-pVDZ -1.722 -0.698 -1.024

aug-cc-pVTZ -1.159 -0.193 -0.966

=10-5 -0.872

=10-7 -0.956

=10-9 -0.956

52

Benzene dimer timings(Sequential Pentium IV 2.4 GHz)

Basis Time/min

k=9 monomer 56

k=9 dimer 200

k=9 total

(D2h)

256

aug-cc-pVTZ total

(D6h)

305

(915 D2h est.)

53

Benzene monomer, dimer and trimer(aug-cc-pVDZ LDA geometry)

• Dimer binding energy -0.96 kcal/mol.

• Trimer -1.67 kcal/mol.

• Single processor times for k=9 energy(energy accurate to about 1e-6).– Monomer 56 minutes– Dimer 200 (3.6x = 21.84)– Trimer 457 (2.3x = (3/2)2.05)

54

Also working• Takeshi Yanai

– Analytic derivatives– Fast (O(N)) Hartree-Fock exchange– TDDFT within Tamm-Damcoff approximation– GGA

• Abelian point group symmetry (D2h + subgroups)

• Thanks also to – So Hirata for guidance with TDDFT– Edo Apra for insights into DFT

55

56

Putting it all together – A path to O(N) exact MP2

• HF provably O(N) to arbitrary finite precision– Based upon the density matrix– Localized orbitals also possible (Bernholc)

• Need an MP2 scheme based upon density matrices

57

The Resolvent has low separation rank• Already known – Almlöf Laplace factorization

1

0

1

2

2

2

1/8 1/8

1

4

1

4

a b i j

a b i j

f i a f

t

a b i j

Mt

ijab a b i j

ijab

t t

e dt

e

ij abE

i j a b

i e i a e a

58

Density matrix form of MP2

0 0

2

1 11 3 2 4 1 3 2 4 12 34 1 2 3 4

2 3 2 3 2 3

( (1) ) ( (2)1/ 41 2

1( | ) 2( | ) ( | )

4

12 ( , ) ( , ) ( , ) ( , )

4

exchange

12 ( , ) ( , ) exchange

4

( , ) f

ijab

H t H

E i a j b i a j b i b j a

r r r r q r r q r r r r dr dr dr dr

V r r V r r dr dr

r r e e

)

1 2

12 3 1 3 1 3 12 1

( , )

1

( , ) ( , ) ( , )

f t r r

q

V r r r r q r r r dr

59

Summary• Multiresolution provides a general framework for

computational chemistry– Accurate and efficient with a very small code– Multiwavelets provide high-order convergence and

accommodate singularities• Familiar orthonormal basis (Legendre polynomials)

– Compression and reconstruction (c.f., FFT)– Fast integral operators (c.f., FMM)

• Separated form for operators and functions– Critical for efficient computation in higher dimension

• Expect speed competitive to Gaussians in near future– Optimal separated forms for kernels, multi-scale non-linear

solver, better implementation• Real impact will be application to many-body models