Benchmarks
These benchmarks measure the cost of one state-advanced nonlinear-flux evaluation across WaveVortexModel transform families and grid sizes. They are intended to help you estimate runtime and memory requirements, compare computers, and reproduce the measurements on your own hardware.
Performance at a glance
The representative comparison uses the nonhydrostatic constant-stratification case with resolution [256 256 65]. Absolute measurements are shown because benchmark scores can obscure the practical time and memory required by a model.
No approved benchmark datasets have been published yet. Issue #140 will add the first two-machine MATLAB results.
Scaling with model size
The plots separate horizontal and vertical scaling. Each point is the median of the retained timing samples. Memory plots report the peak resident memory of the MATLAB or C++ process, including the language runtime and numerical libraries.
No approved scaling datasets have been published yet.
Compare computers
Processor, memory, operating-system, toolchain, and thread information accompany every published dataset. Results from different environments remain machine-dependent measurements rather than universal performance guarantees.
No approved computer results have been published yet.
Run the benchmark yourself
Benchmark tools are authoring utilities and are not installed on the runtime package path. From a clean WaveVortexModel authoring checkout, run:
addpath("Benchmarks")
results = runWaveVortexBenchmark(suites="scaling-standard-v1")
The larger scaling-large-v1 suite can require substantially more memory. The benchmark authoring guide explains suite selection, reference generation, raw artifacts, and normalization for publication.
Methodology and interpretation
The measured operation advances the coefficient state and evaluates nonlinearFlux with ordinary production caches retained. It is not a complete model time step. Published cases must pass their numerical correctness tolerance before their timing can be shown.
Runtime is the median of the recorded post-warmup samples. Peak process memory is the largest observed resident-memory value for the process. Memory above baseline subtracts the fresh-process value measured before constructing the model; it is the better estimate of the additional memory associated with WaveVortexModel, while still including allocator and library behavior.
Only datasets approved in the benchmark catalog are published. Comparisons require the same suite contract, operation, case, domain, resolution, numerical options, random seed, warmup count, and sample count. Missing implementation coverage is reported as unavailable and never as zero runtime or memory.
Downloadable results
The normalized files use the language-neutral published-benchmark-v1 contract. The corresponding raw artifacts retain implementation-specific diagnostics and provenance.
No approved result files have been published yet.