ESPResSo

ESPResSo is a soft-matter molecular-dynamics package for coarse-grained models — polymers, colloids, and charged fluids — with P3M (Particle–Particle Particle–Mesh) electrostatics.

On the Orange Pi RV2 we ran two lines of work:

  1. FFTW axis — double-precision libfftw3.so.3 under one unchanged EESSI pypresso binary via LD_PRELOAD (same pattern as GROMACS / FFTW).
  2. Pair-loop axis — rebuild ESPResSo 4.2.2 with opensolvers patches (slim myconfig, drop std::function indirection, monomorph WCA+P3M loop, SR Ewald table) and compare to stock EESSI.

Benchmark source: opensolvers/benchmarks/espresso. Companion probes: FFTW, GROMACS, ScaFaCoS.

Bottom line: RVV FFTW is noise (~0%) on fair pinned P3M runs — same lesson as small-mesh GROMACS and QE. The real win is source-level pair-loop work: ~1.74× vs EESSI on a dense charged fluid (dense_large, N≈5528). MPI strong scaling adds ~2.8× at 8 ranks on the same case.

FFT axis — scalar vs r5v (run-espresso-fft-ab.sh)

Swaps only libfftw3 under EESSI ESPResSo/4.2.2-foss-2025b. Pin mesh/cao/r_cut/alpha on both sides — auto-tune can pick different meshes if one FFTW is faster during tuning probes.

Fair pinned results (2026-08-30)

case mesh steps scalar r5v speedup
N=512, box=20 18³ 1000 13.23 s 13.26 s 0.998×
N=512, acc=1e−5 24³ 1000 34.72 s 35.56 s 0.976×
N=4096, box=40 32³ 300 57.28 s 57.41 s 0.998×
N=4096 44³ 200 37.15 s 36.84 s 1.008×

Energies bit-match. Even at artificial mesh 48³ (Coulomb ~98% of wall), r5v is still ~0.98×.

perf on EESSI dense_large: libfftw3 ~4% of samples; top symbols are add_non_bonded_pair_force, P3M real-space kernel, and long_range_kernel.

Hot path — where the time goes

Multi-model survey (hotpath_models.py) on dense charged fluids:

model N Coulomb share of MD
lattice512 512 ~88%
dense_large ~5528 ~86%
dilute samples (salt_box50, electrophoresis) — ~0% (not a Coulomb probe)

For dense ± fluids, Coulomb/P3M dominates — but most of that is Espresso’s pair loop and P3M spread/k-space, not FFTW kernels.

Pair-loop axis — optimized build vs EESSI

Built from ESPResSo 4.2.2 source on RV2 (build-espresso-opt.sh → ~/espresso-opt):

patch what it does
slim myconfig 7 features (WCA + P3M) vs EESSI’s 34
ForceKernelRef drop std::function in Coulomb short-range
add_non_bonded_pair_force_p3m monomorph WCA + direct CoulombP3M::pair_force
SR Ewald table 4096-bin lookup replaces per-pair exp
k-space loop pointer locals in long_range_kernel differentiation

dense_large (N≈5528, mesh 32³, 200 steps, 3 reps)

build mean wall vs EESSI
EESSI 4.2.2 46.8 s 1.0×
opt v1 (ForceKernelRef + slim) 31.8 s 1.44×
opt v2 (+ monomorph + SR table) 28.1 s 1.66×
opt v3 (+ k-space patch) 26.8 s 1.74×

Energies match (E_tot / E_coul to ~1e−4). lattice512: 1.35× (EESSI ~8.0 s → opt ~6.0 s).

Levers that did not help further

lever result
Lower cao (4/5) + retune Slower — tuner raises r_cut; pair work dominates assign savings
RVV FFTW on opt build ~0% (scalar 27.5 s vs r5v 28.3 s mean)
Larger Verlet skin Slower — best skin 0.4 for dense fluid

After v2 patches, perf shows add_non_bonded_pair_force and std::function _M_invoke gone from the hot list.

MPI strong scaling

Pinned P3M, scalar FFTW, OMP_NUM_THREADS=1, OpenMPI on 8× X60:

np dense_large wall (200 steps) speedup
1 46.0 s 1.00×
2 26.2 s 1.75×
4 23.6 s 1.95×
8 16.3 s 2.82×

Energies match across ranks. For production Coulomb-heavy runs: MPI + opt binary, not more FFTW tuning.

Prior note (2026-08-21)

Short auto-tuned run (N=512, 200 steps, mesh 18³) once reported ~1.12× RVV FFTW; that did not reproduce under longer fair pins with the current xorconj build.

backend wall speedup
scalar 2.706 s —
r5v 2.425 s 1.12×

Measured: Orange Pi RV2, EESSI dev.eessi.io/riscv 2025.06-001.