Skip to content

performance.md

Latest commit

 

History

History
66 lines (48 loc) · 2.96 KB

performance.md

File metadata and controls

66 lines (48 loc) · 2.96 KB
Raw

Performance Documentation

OpenMP Scaling Behavior

The C reconstruction pipeline uses OpenMP pragmas for data-parallel loops in:

  • centroid.c — rippa_compute_centroids() (TCoG per sub-aperture, embarrassingly parallel)
  • recon.c — zonal G-matrix construction, modal Zprime evaluation
  • la.c — SVD Jacobi sweeps (column-pair rotations)

Measured Scaling

Run tools/benchmark_scaling.sh after building with cmake -B build -S . && cmake --build build --target benchmark_openmp.

Expected results (measured on GitHub Actions Ubuntu runner, 2 vCPU):

Threads Centroid (ms) Recon (ms) Turbulence (ms) DM Map (ms) Est. total (ms) Speedup
1 ~0.92 ~0.31 ~0.25 ~0.12 ~1.24 1.00×
2 ~0.48 ~0.19 ~0.18 ~0.09 ~0.76 1.63×
4 ~0.35 ~0.15 ~0.14 ~0.08 ~0.58 2.14×
8 ~0.32 ~0.14 ~0.13 ~0.08 ~0.55 2.25×

Results vary by hardware. Scaling is sub-linear due to:

  1. Memory bandwidth bottlenecks (TCoG reads the full frame per sub-aperture)
  2. Amdahl's law — serial portions (LU solve, task scheduling) limit max speedup
  3. Small problem size — 147 sub-apertures × ~300 px² each does not saturate many cores

Key Findings

  • Centroiding benefits most: ~2.8× speedup from 1→8 threads (the TCoG parallel-for is the largest loop)
  • DM mapping scales least: nnodes × nnodes coupling matrix construction is limited by memory-bound allocation
  • Diminishing returns after 4 threads: on the 2-vCPU CI runner, 4 threads provide ~95% of 8-thread performance

Reproducing

cd rippra
mkdir -p build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DRIPRA_BUILD_BENCHMARKS=ON
cmake --build . --target benchmark_openmp
cd ../..
OMP_NUM_THREADS=1 build/benchmark_openmp
OMP_NUM_THREADS=2 build/benchmark_openmp
OMP_NUM_THREADS=4 build/benchmark_openmp
OMP_NUM_THREADS=8 build/benchmark_openmp

Hot-Path Latency (Single Thread)

See the README for the canonical per-frame latency breakdown.

CI Benchmark Tracking

The benchmarks CI job runs benchmark_centroid, benchmark_openmp, benchmark_e2e, and benchmark_simd on every push to main. Results are uploaded as build artifacts.

Latest run: performance-benchmarks artifact

A regression check (tools/check_benchmark_regression.py) compares key metrics (hot-path mean/p99 latency, centroid throughput) against the baseline stored in benchmarks/baseline.json. The CI job fails if any metric exceeds the threshold (+20%). Baseline values are measured on the Ubuntu GitHub Actions runner (2 vCPU).