Benchmarks and validation¶

These records preserve numerical and performance evidence for the current runtime. They are engineering measurements rather than universal performance claims: each result is valid for the stated model, case, allocation, software and execution path.

Record index¶

Record	Scope	Principal metrics
Transport	UDS, shared memory, TCP and UCX	Exchanges per second and payload throughput
Runtime	Field preparation, model execution, memory and artefact handling	Latency, throughput, load time and resident memory
OpenFOAM	Solver, MPI and end-to-end coupled comparisons	Wall time, OpenFOAM time and field error

Terminology¶

Term	Meaning
OpenFOAM	Direct execution without FoamNordic
OF/FNO	OpenFOAM launched by Longship without an externally coupled field model
FNO-coupled	OpenFOAM with a resident FoamNordic closure or transform
FNO-native	A symbolic FoamNordic function lowered into the OpenFOAM adapter without ModelHost
Cold	Includes the stated first-load, compilation or activation work
Warm	Excludes the stated warm-up operation

Reporting contract¶

Microbenchmarks use at least seven warm runs unless a section states another repetition count. A value written as \(m \pm d_{\max}\) reports the sample median \(m\) and the largest absolute deviation from that median,

\[ d_{\max}=\max_i |x_i-m|. \]

OpenFOAM comparisons preserve the source case, mesh, decomposition, allocation, time-step policy and output policy between paths. End-to-end wall time includes Longship preparation and shutdown; OpenFOAM execution time is parsed from the solver log. Throughput counts the complete payload stated by the benchmark rather than an inferred hardware bandwidth.

Systems¶

System	Processor	Memory	Software
MacBook	Apple M5, 10 cores (4 performance + 6 efficiency), arm64	16 GB unified	macOS 26.6.2, Apple Clang 21.0.0
Roihu CPU node	2 × AMD EPYC 9965, 384 cores, x86_64	768 or 1,536 GiB	RHEL 9, GCC 15.2.0
Roihu GPU node	4 × NVIDIA GH200, 288 Grace cores, aarch64	480 GiB LPDDR5 + 384 GiB HBM3	RHEL 9, JAX 0.11.1, CUDA 13

Roihu jobs use only the resources stated with each result.

Partition	Nodes	CPU and memory	Time limit
`interactive`	1	Up to 32 CPUs and 64 GiB	36 h
`test`	1–2	Up to 384 CPUs and 744 GiB per node	15 min
`small`	1	Up to 384 CPUs and 1,500 GiB	72 h
`medium`	1–6	Full nodes, 384 CPUs and 744 GiB per node	36 h
`gputest`	1–2	1–4 GH200 GPUs per node; 217 GiB per GPU	15 min

Validation coverage¶

Evidence	Coverage
Native gate	223 Python tests, 17 native tests and `foamnordic doctor` on the MacBook and a Roihu CPU node
Two-node transport	TCP job `1092472` and UCX job `1093015`
Solver-native closures	16-rank Smagorinsky and kEqn job `1095041`
Repeated WALE comparison	Interactive job `1098732`

Return to the documentation index.