QuantumChemistry500

NAKATA Maho (RIKEN)ISHIMURA Kazuya (IMS)

HIRANO Toshiyuki (U of Tokyo)Jeff HAMMOND (Intel)

2014/11/18 Tue Nov 18 @ 12:15pm-1:15pm Room 295

Role of HPC Benchmarks

• Represent important applications. • Based upon simple codes that non-experts

(especially vendors) can optimize them.• Be somewhat orthogonal to each other.• Stress computer hardware in interesting ways.• Allow for objective comparison between

different computing platforms.

Current HPC Benchmarks

• Top500 = HPL: LU factorization (just DGEMM?).• Graph500: Non-numerical benchmark.• HPCG: Conjugate gradient PDE solver for simple

stencil.• HPGMG: Geometric multigride PDE solver.• HPCChallenge: Collection of benchmarks.• STREAM• Scalable Synthetic Compact Applications (SSCA)• DOE Mini-apps• ...

Quantum Chemistry in HPC

• QC/DFT major component of scientific workloads.

Figure courtesy of Richard Gerber (NERSC)

Many QC apps are built by users and un-tracked.

VASP and NWChem build matrix differently. QC500 represents harder way.

What is QuantumChemistry 500?

• Very different properties than existing benchmarks:

– nontrivial load-balancing (irregular, dynamic tasks)

– small- to mid-sized messages (between 8B and 100KB)

– nontrivial to vectorize (short SIMD)

– balance of memory- and compute- intensive

– kernels contain branching

– modest dense linear algebra (not HPL-sized)

• Allows many implementations as long as same numerically.

– Easier entry for novel hardware (VHDL impl???).

• Optimized implementations already exist:NWChem, …, GTFOCK (OpenMP/SSE/AVX), TeraChem (GPU)

What is QuantumChemistry 500?

• Chemistry-specific benchmark targeting most common method(s). Initial target is Hartree-Fock SCF (DFT-like).

• Science-driven, scale-invariant focus: Performance per node/watt/etc…

• Allows different algorithms and software as long as the answer is the same.

• Building upon existing HPC codes for initial data; encourage new optimized code development.

• Exercise hardware using challenging kernels not captured by any existing benchmark.

• Avoid Goodhart’s Law (A machine built just for QC500 will be good at many things...)

Hartree-Fock/SCF/DFT Theory

This is the classic algorithm; variations exist.

● Formation of matrix is irregular.● Matrix elements highly non-

trivial (3+ methods exist).● Diagonalize via GEVP or DMM.

Quantum chemists have implemented many algorithms

in many software packages and yet it is possible to obtain

numerical consistency between codes!

Reference inputs

- Insulin - http://www.pdb.org/pdb/101/motm.do?momID=14 - PDBID: 2HIU - 51 AA

- Antifreeze Proteins - http://www.rcsb.org/pdb/explore.do?structureId=2PNE - PDBID: 2PNE - 70 AA

- Ubiquitin - http://www.pdb.org/pdb/101/motm.do?momID=60 - PDBID: 1UBQ - 76 AA

- HIV-1 Protease - http://www.pdb.org/pdb/101/motm.do?momID=6 - PDBID: 7HVP - 99 AA

2HIU : Insulin

2PNE : antifreeze Protein

Input Specification

• Given data as reference– Atomic coordinates, molecule charge and spin.– Basis set (cc-pVXZ – to allow for a range of problem sizes)– Sample input files for several quantum chemistry packages

to enable data collection by operators.

• Requirements for accuracy/precision of final results.

• Hartree-Fock and DFT (B3LYP)• Other methods under investigation for the future.

Implementations• ACESIII• CFOUR• Dalton• FireFly• GAMESS• Gaussian• GTFock• Molpro• Molcas• NWChem• ORCA• ProteinDF• Psi4• QChem• SMASH• TeraChem• TurboMole• etc.

Submissions must include detailed algorithmic and implementation specification sufficient for reproduction in a different implementation.

Numerical tolerances must be documented.

The best specification includes complete source code.

Reference Results

● We will use GAMESS, NWChem, and ProteinDF to generate reference energy values. >>> They must agree to be a valid reference!

● These codes are free, parallel and widely supported by HPC folks. Involved in many procurements so vendors are familiar.

● Reference codes do not have a lot of approximations by default (no linear-scaling tricks).

Conditions for valid result

• Converged total energy should match our reference value to six (?) decimal places in atomic unit.

• Converged orbital energies should match our reference value to three (?) decimal places in atomic unit.

• These criteria are open for debate. Some may argue for higher accuracy...

Results to be submitted• Elapsed time• Which program package is used & Input file

– Changes from default (e.g., cutoff value)– Details of implementation and algorithms

• Output file– Total energy and orbital energies

• Machine configuration– CPU, memory, network, storage and their peak

(theoretical) performance

• All of above info will be open to the public

TODO until next year

• Forming steering committee of scientific experts with deep HPC knowledge:

• academia/national labs

• industry (IBM, NVIDIA, …)

• Digital presence

• Reference data and validation infrastructure

• The first benchmark results on several supercomputers.