The NWChem benchmark dataset is a set of standard performance test scenarios specifically designed for the NWChem quantum chemistry and molecular simulation software in high-performance computing (HPC) environments. This dataset, generated by the NWChem high-performance computational chemistry software, covers quantum and classical hybrid computational data for biomolecules, nanostructures, and solid-state materials. It includes ground-state and excited-state properties, employs both Gaussian function and plane-wave computational methods, and boasts high parallel scalability from single nodes to thousands of processors. It also supports the analysis of molecular properties and relativistic effects.

The relevant research paper, titled "NWChem: Past, present, and future," was published in 2020 by the Pacific Northwest National Laboratory in collaboration with Lawrence Berkeley National Laboratory, the National Center for Computational Sciences, and other institutions.

Dataset structure

The basic structure of this dataset includes:

• Input scripts and output results: Each benchmark scenario comes with a reproducible computation input file (.nw) and corresponding output logs/timing data (including wall clock, CPU time, speedup).
• Parallel Architecture: Includes runtime configurations with different numbers of processors and parallel modes (MPI + OpenMP or pure MPI) for analyzing performance scalability.
• Multiple chemical methods: Involves computational modules such as molecular dynamics (MD), self-consistent field (SCF), density functional theory (DFT), and second-order perturbation theory (MP2).
• Reference operating environment description: Specify the supercomputing platform used for computing (such as IBM SP2, Cray T3E-900), number of nodes, number of cores, wall-time, etc.
• Images and performance curves: Each scene is accompanied by a performance graph (.gif, .jpg) showing metrics such as speedup ratio, CPU time, and disk usage.
• Download and input reproduction examples: All benchmark tasks can be reproduced by downloading the corresponding input files (such as had_md.nw, siosi3.nw, h2o7.nw, etc.).

Dataset content example

The following are some typical benchmark contents:

• Liquid water systems (Molecular Dynamics):
  Molecular dynamics simulations were performed on systems with 5184, 17496, 41472, and 82000 atoms, using the SPC/E water model with a cutoff radius of 1.8 nm, and run on an IBM SP2.
  The simulation demonstrates the parallel scalability of the liquid water system, showing that each processor needs to handle approximately 100 atoms to achieve good scalability.
  The results show that maintaining a reasonable computational load is crucial for efficiency in massively parallel computing.
• Na⁺/K⁺ Crown ether complex free energy calculation:
  The system containing 6382 atoms was run on IBM SP2 and Cray T3E-900 to calculate the relative free energies of Na⁺ and K⁺ ions in aqueous solution and in their complex with 18-crown-6 crown ether.
  The composite free energy difference was calculated to be approximately 6 ± 4 kJ/mol using the multiconfiguration thermodynamic integration (MCTI) method (another calculation yielded 5 ± 5 kJ/mol), consistent with the experimental value of 7 kJ/mol.
  The results demonstrate the runtime and parallel performance of the system on different platforms.
• Haloalkane Dehalogenase enzyme mimicry:
  The system contains 41,259 atoms and employs an AMBER force field with Ewald (PME) correction and a cutoff radius of 1.0 nm.
  Molecular dynamics simulations of the enzyme in aqueous solution were performed on an IBM SP, with long-range electrostatic energy and force corrections performed using a 64³ grid.
  The results show good parallel acceleration behavior and provide input files that can be run directly (had_md.nw, had.top.gz, had_md.rst.gz).
• 1,2-Dichloroethane droplets:
  A system containing 100,369 atoms was used to simulate the behavior of pollutant droplets.
  Using Paulsen chloroalkane force field parameters and the SPC/E water model, with a cutoff radius of 2.4 nm, the experiment was run on IBM SP and Cray T3E-900, respectively.
  Simulations demonstrate the performance comparison and scalability of 1,2-dichloroethane droplets on different platforms.

Liquid octanol:
The simulation system contains 216,000 atoms, employs the AMBER force field and SPC/E water model, and has a cutoff radius of 2.4 nm.
Running on a Cray T3E-900, the linear acceleration and good scalability of the liquid octanol system under large-scale parallel conditions were demonstrated.

• SCF Performance Testing:
  Semi-direct distributed data flow computing (DDSCF) was performed on an IBM SP (150 MHz node) to examine CPU speedup and disk usage as the number of processor nodes increases.
  The results show that disk usage increases proportionally to available resources, verifying the scalability of the SCF module in a distributed environment.
• DFT benchmark (SIOSI3/6/7):
  LDA calculations were performed on three zeolite fragment systems (containing 347, 1687, and 3554 basis functions, respectively) to evaluate the scalability of the density functional theory module.
  Provide input files (siosi3.nw, siosi6.nw, siosi7.nw) for reproducible experiments, requiring computations to be performed entirely in memory (in-core).
  The memory usage can be verified by searching for the keyword "in-core" in the output file, and the calculation results show good multiprocessor parallel acceleration.
• MP2 gradient calculation:
  MP2 gradient calculations were performed on the (H₂O)₇ molecule and potassium crown ether system to analyze the CPU time distribution under different numbers of processors.
  The calculations were performed on an IBM SP (120 MHz node), demonstrating the proportion of computation time for each part.
  A reproducible input file, h2o7.nw, is provided for further performance testing.

Usage Tips

• For performance comparisons, please record the number of nodes/cores, wall-time, and parallel configuration (MPI×OpenMP) to ensure a fair comparison.
• Before running, please refer to the input file and README document corresponding to each benchmark scenario to ensure that the input is consistent and the parameters are complete.
• The benchmark can be used for:
• New hardware platform validation: NWChem performance evaluation under environments such as GPU, hybrid systems, and accelerators.
• Parallel optimization evaluation: performance impact of different compilation options, MPI communication model, and memory scheduling.
• Comparative study of software: performance comparison with other quantum chemistry programs (such as Gaussian, CP2K, ORCA) on similar systems.