Parallelism & Drivers

One parallelism level, chosen deliberately

Blaze parallelizes whole solve jobs, nothing else. A parameter sweep expands into independent jobs; rayon runs them data-parallel on a dedicated thread pool; the only shared state is the FFT plan cache and a mutex-guarded output writer. Inside one job, everything is single-threaded.

Two decisions the board makes visible:

• The k-point loop is sequential by design (board). Each k-point warm-starts from the previous solution, with the subspace gauge-aligned by polar-decomposition parallel transport. That chaining collapses iteration counts (see the eigensolver) but makes k-points a dependency chain, so the parallel unit has to be the job.
• Dense-algebra parallelism is switched OFF (Par::Seq, the badge on the GEMM node). At production grid sizes (≤64²), thread synchronization inside the GEMM made it 5 to 10 times slower than sequential. There is not enough work per GEMM to amortize a fork. The parallelism budget is spent where the work is: across jobs.

Both schedules below use the same four cores. The top machine gives each worker a whole solve; the bottom machine splits one solve’s GEMMs across threads and pays a sync barrier at the end of every micro-phase. Watch the job counters:

Adaptive threading

Job wall-times vary with sweep parameters, so a fixed thread count is rarely optimal. The adaptive thread manager watches live job timings and tunes the pool. With --threads -1 the driver manages itself.

The driver around one job

Per job, the band-structure driver owns the k-loop: clone the dielectric, rebuild the k+G tables, warm-start the block, solve, hand eigenvalues to band tracking. Band indices are matched across k by overlap (polar decomposition plus Hungarian assignment) so bands stay physically continuous through crossings instead of being sorted by value.

Streaming is part of the same story: results leave through a mutex-guarded CSV writer, a crossbeam channel to Python, or per-k callbacks to the browser, so a ten-thousand-job sweep never accumulates in memory.

→ Continue with Interfaces for how those results cross the language boundaries.