Skip to content

performance.md

Latest commit

 

History

History
58 lines (38 loc) · 2.85 KB

performance.md

File metadata and controls

58 lines (38 loc) · 2.85 KB

Performance and Benchmarks

This document describes how to run benchmarks, interpret results (cells/second, memory), and optional tuning (XLA, mixed precision).

Running Benchmarks

Benchmark tests live under tests/performance/. Run them with pytest:

pytest tests/performance/ -v
  • test_solver_speed_small_grid: Runs a small grid for 500 steps and reports cells/second. Baseline: > 10k cells/s (relaxed for CI).
  • test_solver_scaling_with_grid_size: Checks that runtime scales roughly with grid size.
  • test_memory_usage: Measures RSS delta for a medium grid (requires psutil).

Interpreting Results

  • Cells per second = (Nx × Ny × Nz) × n_steps / elapsed_seconds. Higher is better. Typical range: 10k–10M+ depending on hardware and grid size.
  • Memory: Expect on the order of 6 field arrays × 4 bytes × cell count, plus solver/CPML overhead. The memory test compares used RSS to a rough upper bound.

Device and Optional Tuning

Use emsim.device.setup_device() before creating the grid/solver:

from emsim.device import setup_device

# GPU if available, else CPU
setup_device("auto")

# Optional: XLA JIT (can speed up GPU/CPU)
setup_device("auto", enable_xla=True)

# Optional: mixed precision on GPU (higher throughput, slight numerical difference)
setup_device("auto", enable_xla=True, mixed_precision=True)
  • enable_xla: Turns on TensorFlow XLA JIT; often gives 2–5× speedup when applicable.
  • mixed_precision: Sets Keras policy to mixed_float16 on GPU for faster ops; use with care for accuracy-sensitive runs.

Non-uniform (stretched) grid efficiency

When only part of the domain needs fine resolution, a non-uniform grid (variable dx, dy, dz per direction) can reduce total cell count and runtime compared to a globally fine uniform grid. To evaluate gains for your setup:

  • Run the benchmark and inspect tables/plots in Tutorial/02_grid/: 
    python Tutorial/02_grid/run_efficiency_benchmark.py
  • See Tutorial/02_grid/results.md and results.csv for metrics (cells, time, cells/s) comparing uniform vs non-uniform for a representative scenario (local refinement in the centre). Use the same idea to add your own scenarios and plots.

Expect noticeable gain when a small fraction of the domain requires fine resolution; gain is marginal when most of the domain is already fine.

Implementation Notes

  • @tf.function: The core field updates update_H and update_E in emsim.fdtd.fields are compiled with @tf.function(reduce_retracing=True) to reduce Python overhead.
  • set_region: Material grid updates use tensor indices (e.g. tf.meshgrid + tensor_scatter_nd_update) instead of Python loops for large regions.
  • Port recording: Modal and lumped ports record into tf.TensorArray; conversion to NumPy happens only at the end when building S-parameters or results.