OutWit.Controller.CalculiX 1.1.2

OutWit.Controller.CalculiX

Runs complete CalculiX (ccx) decks on WitCloud compute nodes. One solve = one whole deck on one node — no decomposition, no model conversion; deck support is whatever the pinned ccx solves. The win is throughput across many independent solves, which is exactly what the companion OutWit.Controller.Sweep orchestrates into parameter studies.

Activity

The task rides as one envelope (CcxTaskData): deck blob id, explicit node/element counts (work estimation never opens the blob), thread policy and the extraction request. Results return in completion order — consumers map by VariantIndex, never positionally.

Bundled solver

The module carries pinned ccx 2.22 builds for win-x64, linux-x64 and osx-arm64 as controller data assets, produced and mirrored by OmnibusCloud/CalculiX. Nodes need no preinstalled software - the macOS kit carries the GCC runtime (libgfortran/libgomp/libquadmath/libgcc_s) beside ccx, referenced through @loader_path; ccx-v2.22-1 had linked them by absolute Homebrew paths and every variant on the first Apple Silicon node died with "dyld: Library not loaded" (exit 134) - ccx-v2.22-3, controller 0.1.8. The Windows kit is built on MSYS2 winpthreads 13.0.0.r488: the 14.x runtime turns every uncontended mutex into a kernel wait, and libgfortran takes one per Fortran I/O statement, so ccx-v2.22-3 ran the reference cube 1.8x slower than upstream's own Windows build and Windows workers were rated below Linux ones by the same factor - ccx-v2.22-4, controller 1.0.1, results byte-identical. CalculiX is GPL-2.0: the asset kit ships the license text and the written source offer, and the corresponding source is publicly mirrored in that repository's releases.

Node benchmark: Ccx.Solve is ranked by the reference deck (a 40³-node static cube, 64 000 nodes, generated by CcxReferenceDeck so it is the same on every machine) in unit ccx-static@ref40-v1. ref20-v2 (controller 1.0.3 and earlier) used a 20³ cube that solves in under a second, where process start and one core decide the time: a 2-vCPU virtual machine tied a 16-core desktop and the fleet's strongest machines ranked last. The 40-cube takes 8-25 s and lets the cores count (a Ryzen 9 5950X 8.4 s, a 6-core i7-8700K 10.9 s, a 2-vCPU VM 17.8 s), so a benchmark costs a node half a minute to a minute once per controller install. Solves run with at most 16 OpenMP threads (CcxProcessRunner.MAX_DEFAULT_THREADS): PARDISO saturates by then and SMT siblings only cost time. The first solve is an untimed warm-up - it is the first start of the freshly installed kit, and its cold page cache and antivirus scan once rated a Ryzen 9 5950X at 1.36 s against the 0.75 s it takes from then on; the rate is 1 / the median of three to five timed solves (1.5 s target), so one stall does not move it and a busy machine is still rated as busy. ref20-v1 (controller 1.0.1 and earlier) timed one cold solve. The Custom bag carries the run times (runs_s, median_s, warmup_s) and the solved maximum displacement (checksum).

ccx is started without a console window (CreateNoWindow). The worker client is a windowed application with no console of its own, so before controller 1.0.3 every ccx.exe it launched got a fresh console, which cost about 0.45 s per start on Windows 11: the reference cube took 1.13-1.23 s from the client against 0.69-0.76 s with the flag, and every sweep variant on a Windows node paid it.

Determinism note: ccx with OpenMP is not bitwise-reproducible across thread counts, and the three platform builds add last-digit variation — results are stable to engineering tolerance, and the controller's tests assert tolerance, not bits.

SPOOLES runs with one equation-solver thread (CCX_NPROC_EQUATION_SOLVER=1, CcxEquationSolver). The kits link SPOOLES multithreaded, and multithreaded SPOOLES returns a wrong field now and then with exit 0: a 14 k-node bracket solved 60 times in concurrent pairs came back different in 2 runs (by up to 0.6 %), and a sweep variant on an arm64 Mac was 10 % off. PARDISO and single-threaded SPOOLES gave the same answer every time. A solve goes through SPOOLES when a step card says SOLVER=SPOOLES, and always on macOS, whose kit has no PARDISO (ccx's default where it is linked); assembly and stress recovery keep their OpenMP threads, and the single-threaded factorization cost about 15 % on that bracket. The benchmark's reference solve follows the same rule.

Dependencies

Variables (module dependency). The shared data types live in OutWit.Controller.CalculiX.Model, consumed by this controller, by the Sweep orchestration controller (a CalculiX study carries the extraction request; its manifest rows carry the CcxResultData verbatim) and by client applications. The sweep's own types live in OutWit.Controller.Sweep.Model.