Neighbor List Benchmark Results#

This directory contains pre-computed benchmark results for different neighbor list algorithms on various GPU hardware.

File Naming Convention#

Results are stored in CSV files with the following naming pattern:

neighbor_list_benchmark_<method>_<gpu_sku>.csv

Where:

<method>: The neighbor list algorithm (naive, cell_list, batch_naive, batch_cell_list)
<gpu_sku>: GPU identifier (e.g., rtx_a3000_laptop_gpu, a100_sxm4_80gb)

Running Benchmarks#

To generate new benchmark results:

cd benchmarks/neighborlist
python benchmark_neighborlist.py --config benchmark_config.yaml --output-dir ../../docs/benchmarks/benchmark_results

Command Line Options#

--config: Path to YAML configuration file (required)
--output-dir: Output directory for CSV files (default: ../../docs/benchmarks/benchmark_results)
--methods: Specific methods to benchmark (e.g., --methods naive cell_list)
--gpu-sku: Override GPU SKU name for output files

Examples#

Run all benchmarks:

python benchmark_neighborlist.py --config benchmark_config.yaml

Run only specific methods:

python benchmark_neighborlist.py --config benchmark_config.yaml --methods naive cell_list

Override GPU SKU name:

python benchmark_neighborlist.py --config benchmark_config.yaml --gpu-sku custom_gpu_name

CSV Format#

Each CSV file contains the following columns:

method: Algorithm name
total_atoms: Total number of atoms in the system
atoms_per_system: Atoms per system (for batch methods)
total_neighbors: Total number of neighbor pairs found
batch_size: Batch size (1 for single-system methods)
median_time_us: Median execution time in microseconds
success: Whether the benchmark completed successfully (optional)
error: Error message if benchmark failed (optional)
error_type: Error type (e.g., “OOM”, “Timeout”) if failed (optional)

Visualization#

The Sphinx documentation automatically discovers and visualizes all CSV files in this directory. See the benchmarks section of the documentation for interactive plots and comparisons.