Eve library

Kalika MD performance benchmarks

Other

Revision 1

On this page

All timings use optimized Rust (cargo run --release), deterministic fixtures, and std::hint::black_box. CSV output includes the phase-120 target and whether the local measurement passes it. Run one benchmark process at a time.

Before a full size matrix, follow the repository resource-safety gate. The CI smoke profile is intentionally small and verifies execution, not performance:

sh
cargo run -p md-benchmarks --release --bin classical-md-benchmark -- --smoke

The full classical matrix measures 1K, 10K, 100K, and 1M atom LJ force steps with a retained Verlet list and PME steps with linked-cell real-space work. Override repetitions with KALIKA_BENCH_ITERATIONS; use KALIKA_LJ_SIZES or KALIKA_PME_SIZES to select an explicit subset after reviewing host resources.

The comparator column records whether a LAMMPS serial executable (lmp or lammps) is available. It does not substitute or estimate an external result.

Baseline: 2026-07-13#

The first full single-core run on the 15 GiB Linux validation host used one timed force step per size (release/LTO build):

kernel atoms atom-steps/s phase target result
LJ + retained Verlet list 1,000 2,439,310 1,000,000 pass
LJ + retained Verlet list 10,000 1,586,952 1,000,000 pass
LJ + retained Verlet list 100,000 2,210,653 1,000,000 pass
LJ + retained Verlet list 1,000,000 2,129,851 1,000,000 pass
PME + linked-cell real space 1,000 337,951 100,000 pass
PME + linked-cell real space 10,000 340,591 100,000 pass
PME + linked-cell real space 100,000 309,838 100,000 pass
PME + linked-cell real space 1,000,000 315,284 100,000 pass

Neither lmp nor lammps was installed, so no same-host LAMMPS number is reported. The local results retain the original phase targets verbatim; the comparator is never estimated or substituted.

Native versus WebAssembly bindings#

scripts/build-bindings.sh builds the exact same Rust entry point as a napi-rs Node addon, a wasm-bindgen Node module, and a browser-targeted WASM module. It compares neighbor-list construction, retained-list LJ forces, and PME in Node and in a real Playwright-launched Chromium page. Each result includes a checksum; the comparators abort before reporting ratios if native and WASM disagree.

Use KALIKA_BINDING_SMOKE=1 for the small build/ABI gate. The default profile uses 10K atoms for neighbor/force kernels and 1K atoms for the current PME reference kernel. The phase gate is wasm_over_native <= 2.

The 2026-07-13 same-host browser release run produced:

kernel native (s) browser WASM (s) WASM/native result
neighbor list, 10K × 5 0.4895 0.5663 1.157 pass
LJ force, 10K × 5 0.1258 0.1264 1.005 pass
PME, 1K × 3 0.0134 0.0162 1.210 pass

The smoke profile is an ABI/correctness gate only: its very short calls can be dominated by module-boundary and timer overhead, so its ratios are not used as performance acceptance values.

ReaxFF#

The reactive benchmark parses the vendored canonical C/H/O force field and profiles neutral methane boxes at 500, 2,000, and 10,000 atoms. The evaluator's opt-in profiler reports QEq, raw/corrected bond order, bonded corrections, and nonbonded wall time without changing its normal WASM-safe path.

sh
cargo run -p md-benchmarks --release --bin reaxff-benchmark
cargo run -p md-benchmarks --release --bin reaxff-benchmark -- --smoke

The CSV retains an explicit LAMMPS availability field. This host has no lmp or lammps, so the requested ≤3× reax/c comparison cannot be computed here.

The 2026-07-13 one-step release profile measured:

atoms steps/s QEq (s) bond order (s) bonded corrections (s) nonbonded (s) total (s)
500 4.3013 0.2137 0.0085 0.0032 0.0071 0.2325
2,000 0.08046 12.1518 0.1157 0.0509 0.1093 12.4279
10,000 0.002243 438.7440 2.8292 1.5039 2.7277 445.8049

This baseline exposes QEq as the scaling bottleneck (98.4% of the 10K step), which is actionable performance evidence rather than a hidden or extrapolated result.

Trajectory analysis#

The trajectory runner uses 10K and 100K atom, 1,000-frame deterministic trajectories. MSD streams every frame without retaining 2.4 GiB of coordinates. RDF samples ten frames and CNA/Voronoi sample one frame by default, matching normal structural-analysis cadence while still reporting the declared 1,000 frame trajectory and exact sampled-frame count. All neighbor discovery uses the production linked-cell implementation.

sh
cargo run -p md-benchmarks --release --bin trajectory-benchmark
cargo run -p md-benchmarks --release --bin trajectory-benchmark -- --smoke

KALIKA_RDF_FRAMES and KALIKA_STRUCTURAL_FRAMES can expand the sampling matrix. Acceptance gates are one million RDF pairs/s and 100,000 CNA atoms/s. The comparator column reports OVITO availability; OVITO is absent on this host, so no external Python API timing is inferred.

The 2026-07-13 release profile measured:

kernel atoms sampled frames throughput target result
MSD 10,000 1,000 16.37M atoms/s — measured
RDF 10,000 10 4.24M pairs/s 1M pass
CNA 10,000 1 829,627 atoms/s 100K pass
Voronoi 10,000 1 207,865 atoms/s — measured
MSD 100,000 1,000 16.16M atoms/s — measured
RDF 100,000 10 4.37M pairs/s 1M pass
CNA 100,000 1 928,597 atoms/s 100K pass
Voronoi 100,000 1 216,792 atoms/s — measured