SoftwareNTChem

Performance data (1)

Overview

RI-MP2 calculations for the world-record-scale nanographene dimer (C₁₅₀H₃₀)₂ could be computed in 12 minutes and 39 seconds using the K computer with 80,199 nodes and 641,592 cores at the effective performance of 3.1 PFLOPS(30% of the peak performance).

Calculation target

• Energy calculation for a two-layer nanographene dimer, (C₁₅₀H₃₀)₂
• Size: 360 atoms, 9,840 atomic orbitals

Calculation conditions

• Algorithm: RI-MP2 (Resolution of Identity – Second-order Møller–Plesset perturbation method)
• Basis function: cc-pVTZ
• Libraries used: Fujitsu SSLII (BLAS，LAPACK，ScaLAPACK)，SMASH1.2，Libint 1.1.5

Computer

• Machine: The K computer
• Specifications

Performance optimization

• Optimization of calculation: DGEMM(Level 3 of BLAS) was applied by rewriting the four centers two-electron integrals (O(N5)),used in the MP2 perturbation calculation, into the matrix-matrix product,
• Optimization of communication: Implementation of hybrid parallelization with MPI (Message Passing Interface)  and OpenMP (Open Multi-Processing). In particular, for MPI Parallel, we propose a new parallel algorithm (biaxial parallel method applied to integral calculations) that can provide good scaling for all the K computer nodes. In this way, large-scale electronic correlation calculations (RI-MP2/cc-pVTZ) were achieved.

Performance information

• Maximum used nodes: 80,199 nodes and 641,592 cores
• Execution time: 12 minutes 39 seconds
• Effective performance: 3.1 PFLOPS
• Ratio to peak performance: 30%
• Parallelization ratio: 99.99968%
• Scalability: It is shown in the figure below.

Reference information:

• When 8,911 nodes are used, the calculation can be executed in 45 minutes with the ratio to peak performance of 62%.
• (C₉₆H₂₄)₂(240 atoms and 6432 orbitals): With 32,768 cores, the calculations can be executed in 11 minutes and 57 seconds.
• C₂₄₀ fullerene (240 atoms and 7,600 orbitals): The calculations can be executed in 13 minutes.

Performance data (2)

Overview

Molecule pincette (buckycatcher): RI-MP2 energy differential calculations for C₆₀@C₆₀H₂₈ using biaxial paralleled NTChem were computed in 15 minutes and 55 seconds with high parallel calculation using 65,536 cores.

Calculation target

• Molecular pincette (buckycatcher): Energy differential calculation for C₆₀@C₆₀H₂₈
• Size: 148 atoms and 3,884 atomic orbitals

 Calculation conditions

• Algorithm: RI-MP2 (Resolution of Identity – Second-order Møller–Plesset perturbation method) gradient
• Basis function: def2-TZVP
• Libraries used: Fujitsu SSLII (BLAS，LAPACK，ScaLAPACK)，SMASH1.2，Libint 1.1.5

Computer

• Machine: the K computer
• Specifications

Performance optimization

• Optimization of calculation: DGEMM(Level 3 of BLAS) was applied by rewriting the four centers two-electron integrals (O(N5)),used in the MP2 perturbation calculation, into the matrix-matrix product,
• Optimization of communication: Implementation of hybrid parallelization with MPI (Message Passing Interface)  and OpenMP (Open Multi-Processing). In particular, for MPI Parallel, we propose a new parallel algorithm (biaxial parallel method applied to integral calculations) that can provide good scaling for all the K computer nodes. In this way, large-scale electronic correlation calculations (RI-MP2/cc-pVTZ) were achieved.

Performance information

• Maximum used nodes: 8,192 nodes and 65,536 cores
• Execution time: 15 minutes and 55 seconds
•  Parallelization ratio: 99.9981%
•  Scalability: It is shown in the figure below.

Performance data (3)

Overview

To carry out a Quantum Monte Carlo method calculation for HCl molecules, the diffusion Monte Carlo method was applied using 8,192 cores and a parallelization efficiency of 84% was achieved.

Calculation target

• Calculation of the statistical physical quantity of HCl molecules using a Quantum Monte Carlo method.
• Size: 2 atoms, 819,200 Variational Monte Carlo (VMC) samples, and 409,600 Diffusion Monte Carlo (DMC) samples

Calculation conditions

• – Algorithm: VMC method, DMC method
• – Basis function: The Restricted Hartree-Fock (RHF) wave function with the cc-pVTZ-DK basis set
• – Libraries used: Fujitsu SSLII (BLAS, LAPACK, ScaLAPACK), Libint 1.1.5

Computer

• Machine: the K computer
• Specifications

Performance optimization

• Optimization of calculation: Molecular Orbital (MO)calculations were simultaneously performed and DGEMM(level 3 BLAS) was applied.
• Optimization of communication: MPI parallelized by simple distributing walkers (parallel of ensemble).

Performance information

• Maximum used nodes: 1,024 nodes and 8,192 cores
• Parallelization ratio: 99.9977%
• Scalability: It is shown in the figure below.

Reference information

• Parallelization efficiency of 90% for OsO_4, using the VMC method with 25,600 cores.