Performance

This page compares the speed of ras-commander's public methods with the native engine and with h5py, measured for the 0.1.0 release.

The first chart shows the native speed-up for each benchmarked method; the second shows how the cell-polygon gain changes with mesh size.

Results

Repeat calls (warm) on the heaviest test file for each method, medians of three fresh processes. Native is the default auto mode; native mode gave the same timings within noise. Speed-up is h5py time divided by native time.

Method	Result rows	h5py	Native	Speed-up
HdfMesh.get_mesh_cell_polygons	340,246	17,145 ms	939 ms	18.3×
HdfResultsMesh.get_mesh_cells_timeseries (water surface)	1	1,372 ms	184 ms	7.5×
HdfResultsMesh.get_mesh_timeseries (water surface)	193	1,415 ms	213 ms	6.6×
HdfResultsMesh.get_mesh_timeseries (face velocity)	865	1,486 ms	291 ms	5.1×
HdfFluvialPluvial.calculate_fluvial_pluvial_boundary	4,569	10,452 ms	6,831 ms	1.5×
HdfMesh.get_mesh_cell_points	352,141	262 ms	251 ms	1.0×
HdfResultsMesh.get_mesh_max_ws	340,248	489 ms	493 ms	1.0×
HdfResultsMesh.get_mesh_summary (maximum face velocity)	177,301	1,732 ms	1,745 ms	1.0×
HdfResultsMesh.get_mesh_max_face_v	705,332	6,974 ms	7,039 ms	1.0×
HdfXsec.get_cross_sections	1,769	81 ms	84 ms	1.0×
HdfResultsPlan.get_volume_accounting	1	1.1 ms	1.2 ms	0.9×
HdfBase.get_unsteady_timestamps	865	5.6 ms	6.3 ms	0.9×
HdfBndry.get_bc_lines	13	12 ms	15 ms	0.8×
HdfResultsMesh.get_boundary_conditions_timeseries	4	51 ms	62 ms	0.8×
HdfStruc.get_structures	835	34 ms	42 ms	0.8×
HdfResultsPlan.get_runtime_data	1	2.7 ms	3.5 ms	0.8×
HdfMesh.get_mesh_area_attributes	0	3.5 ms	6.8 ms	0.5×
HdfResultsXsec.get_ref_lines_timeseries	11	13 ms	27 ms	0.5×
HdfResultsXsec.get_xsec_timeseries	5	13 ms	29 ms	0.4×

• Across all 50 method × file cases (geometric mean), the native engine is 1.07× faster on the first call in a process and 1.30× faster on repeat calls.
• Every case returned identical data in h5py, auto and native mode: the benchmark hashed the normalized result (values, dtypes, columns, index, CRS and geometry) of every run. There were 0 errors in 600 runs.
• Results with few rows can still be expensive: get_mesh_cells_timeseries returns one cell's series but must locate it in a large compressed dataset.

Benchmark method

• Harness: each measurement runs in a fresh Python process. The mode is chosen only through RAS_COMMANDER_FAST_HDF (h5py, auto, native). The three modes are interleaved (h5py, auto, native, native, auto, h5py, ...) so drift affects them equally. Each process records the first call and two repeat calls.
• Cases: 19 public methods on 7 HEC-RAS result files, from a 3,314-cell mesh to a mesh of about 340,000 cells and 705,000 faces.
• Machine: Windows 11, Intel Core i7-11700K (8 logical CPUs available to the VM), 48 GB RAM, Python 3.13, h5py 3.14 (HDF5 1.14.6), numpy 2.3, pandas 2.3. No other heavy process ran during the 31-minute run (CPU monitored every 30 s).
• Cache: the operating system's file cache was warm for every column; "first call" means a fresh process, not a cold disk. Cold-disk reads were not measured.

Reading the numbers

• Large gains come from work the engine does in Rust that ras-commander used to do in Python: assembling 2D cell polygons from face topology, and selecting rows or cells from compressed time-series chunks without decompressing more than needed.
• No gain for whole-array summaries: both paths decompress the same chunks with zlib, and that dominates.
• Small losses on small reads come from a fixed cost per native read (opening the session, proving the layout). They are a few milliseconds per call and are the main target for future optimization.