Historical task source. The Eve task board holds current task state.
Date: 2026-07-15
Actionable gap closed#
The prior 157.13.3 implementation was only an Amdahl/Gustafson prediction.
neith-sim-farm::multi_gpu now provides the executable grid path:
- exact rectangular decomposition and padded halo materialisation;
- concurrent rank dispatch through independently owned
GridDeviceworkers; - deterministic owned-cell assembly after every stencil iteration;
- a real optional wgpu backend that enumerates through one native compute API (preventing duplicate API views), requests one device/queue per physical adapter—including identical-model adapters—and dispatches the Jacobi compute shader on every rank;
- default-build WGSL parse/semantic validation, multi-rank CPU parity, explicit concurrency proof, halo-boundary checks, and on-box wgpu-vs-CPU parity.
ScalingMeasurement and sim-validation::measured_multi_gpu_gate accept median
wall-clock samples, calculate actual speedup and efficiency, and reject
single-device evidence. This completes every locally implementable part of the
157.14.10 gate without misrepresenting a prediction as a measurement.
Hardware boundary#
The host exposes one Red Hat Virtio 1.0 GPU PCI display device, one DRM render
node, and one wgpu compute adapter: llvmpipe (LLVM 20.1.2, 256 bits) (Gl).
There is no nvidia-smi executable or second physical adapter. The wgpu runtime
parity test succeeds through llvmpipe, but one software adapter cannot produce a
multi-GPU efficiency measurement. Therefore 157.14.10 correctly remains
unchecked until the same measured gate runs on at least two distinct physical
GPUs and reports efficiency of at least 0.80.
Verification#
The final command outputs and commit are the authoritative record. Applicable gates are:
- all fourteen phase-157 crate test suites;
neith-sim-farmdefault and--features gputests;neith-sim-farmandneith-sim-validationClippy with warnings denied;- targeted release builds for both changed crates;
- formatting of every changed Rust source file.