CFD Corrector Performance Notes

This document captures the current performance posture of the CFD corrector, how to measure it, and how to tune it for different use cases.

Measurement

• Harness: dev/dev/cfd_perf.clj
• Run (from repo root):

  clj -X dev.cfd-perf/run :sizes '[32 64 100]' :runs 2 :iterations 20 :compare-parallel? true
  

• Options:
  • :sizes — grid edge lengths (nx = ny = nz)
  • :runs — samples per size
  • :iterations — CG iterations in the Poisson solve (runtime scales ~linearly)
  • :compare-parallel? — run both serial and :parallel? true
  • :dimensions — 2 or 3 (default 3)

Current numbers (runs=2, iterations=20, 3D, Apple M4 Max 16c/16t, 128 GB)

• 32³: serial ~246 ms; parallel ~141 ms (Poisson ~94 ms)
• 64³: serial ~1093 ms; parallel ~423 ms (Poisson ~151 ms)
• 100³: serial ~4351 ms; parallel ~1668 ms (Poisson ~624 ms)

Parallel is opt-in ({:parallel? true}) and auto-disables on tiny problems (n <= 150k interior cells).

Knobs and trade-offs

• Iterations: Fewer CG iterations reduce time linearly; trade accuracy. For interactive use, try 10–20 iterations; for higher fidelity, 40+.
• Grid size: Time scales ~O(n³). Downsample grids for higher Hz.
• Parallel: Enable :parallel? true in corrector/correct for 64³ and larger. Small grids may not benefit.
• Surrounding work: The corrector still does divergence/gradient/boundary passes on maps; heavy map-to-array conversions remain. Further gains would come from flattening more of the pipeline or SIMD/GPU paths.

“Good enough” guidance by use case

• Robotics / autonomy planning (5–20 Hz budget):
  • 32³ at 10–20 iterations can approach ~5–10 Hz in parallel.
  • 64³ parallel ~0.4 s → ~2–3 Hz; reduce iterations or grid to hit higher rates.
• Interactive tuning / UI (1–5 Hz): 64³–100³ parallel is acceptable for human-in-loop iteration.
• Real-time (60 Hz): Not feasible on pure Clojure/Java arrays; would require SIMD/accelerated native kernels or coarser grids with very few iterations.

Tuning checklist

• Reduce :iterations for faster but less accurate projections.
• Set :parallel? true for grids >= ~64³.
• Consider coarser grids or subregion solves for higher update rates.
• Use the perf harness before/after changes to quantify impact.