Runtime and model execution¶

These measurements cover work outside the CFD equations: field transforms, model-input preparation, inference, artefact handling and resident memory.

Measurement index¶

Throughput is reported as processed rows or source bytes per measured second. Unless a section states otherwise, central values and deviations follow the benchmark reporting contract.

Field transform¶

Two hundred in-place fused transforms were applied to one million three-component float64 cells. Seven warm runs were reduced by their median.

The final field range was [1, 7.31602] in every run.

Model-input packing¶

Two hundred packs converted three one-million-cell float64 scalar fields into one interleaved float32 tensor. The comparison is an idealised two-pass cast and interleave.

The MacBook used seven runs. Roihu job 1090004 used one core from a small allocation and measured 21 runs after one warm-up. Checksums were identical.

Equinox model execution¶

An Equinox MLP with ten inputs, three 128-wide hidden layers and one output ran on 262,144 rows per call. Each platform completed one unmeasured compilation and warm-up call followed by 20 measured inference calls.

Cold activation includes FNOM loading, evaluator construction, JIT compilation and first inference, but excludes Python and JAX import. Warm inference includes NumPy-to-JAX input and contiguous NumPy output. The MacBook used JAX 0.11.0; Roihu job 1093649 used JAX 0.11.1 on one GH200. The script is tools/benchmarks/equinox_runtime.py.

Joblib and ONNX model execution¶

Roihu job 1091323 used one thread on a test node. Both backends used the same scikit-learn MLP with ten inputs, three 128-wide hidden layers and one output.

Cold load used 11 repetitions.

Warm inference used 21 repetitions per batch. ONNX used single-threaded ONNX Runtime 1.29.0 in Python; the result excludes Fjord transport and process startup.

Compiled estimator crossover¶

Roihu job 1091731 compared compiled and Joblib FNOM on one thread using an ExtraTrees regressor with 32 trees, maximum depth 10 and ten inputs.

The compiled initial load includes native compilation: its median compiler time was 7.399e-1 s over seven fresh caches. Eleven subsequent cache hits loaded in 2.437e-3 ± 1.233e-3 s. Joblib load also used 11 repetitions.

Warm inference used 21 repetitions per batch. Compiled execution is faster for the two smaller batches; the paths are similar at 4,096 and 262,144 rows.

Resident ModelHost memory¶

The production ModelHost ran a deterministic affine kernel over SHM. Each run used 20 exchanges of 1,024 rows, three of 1,048,576 rows, then 20 small exchanges. Large float64 inputs and outputs occupied 80 MiB and 8 MiB.

Each platform used seven independently started processes. The table reports medians and maximum observed deviations.

Internal accounting was identical on both platforms. MacBook RSS remained at its high-water mark; Roihu returned to 50.160 MiB after the final small exchanges. Both retained only the 1,024-row active workspace. Roihu job 1093451 used one test node.

Runtime generation turnover¶

Each of seven MacBook stress series published 256 immutable runtime generations with 64 KiB payloads. Every publication atomically replaced the selector and ran garbage collection. One generation stayed pinned; another pin moved every 16 publications. After both pins were released, only the current and most recent generations survived, matching retention policy.

The 1,792 publications preserved the current generation and both pinned generations. Released pins were collected. The script is tools/benchmarks/generation_turnover.py.