Performance

Commit-by-commit benchmark history for JustRelax's pseudo-transient and DYREL Stokes solvers and its thermal diffusion solver on the CPU and GPU backends. Main-branch results are recorded by Buildkite on the benchmark-data branch; the harness, cases, and performance model are described in benchmarking/README.md.

Runtime is measured. The effective memory throughput and the effective FLOP rate use the versioned algorithmic model recorded with each result. The roofline compares them against a STREAM triad and an FMA-chain kernel measured on the same hardware. The nonlinear Stokes and DYREL cases, whose cost per cell depends on the rheology, record runtime only.

BackendHistory metricMedian runtimeEffective GFLOP/sT_effAllocationsDownload JSON

Commit history

Latest roofline position

Measured time with modeled FLOPs and bytes.

Peak compute (GFLOP/s)Peak bandwidth (GB/s)

Latest results

Change is relative to the mean median runtime of up to five earlier runs of the same benchmark on the same device.