Tuning
Distributed solves and factorizations
The A \ b, cuNumeric.solve, cholesky, and qr APIs select distributed tasks automatically. Selection follows cuPyNumeric 26.06:
| Operation | cuSolverMp cutoff | Internal block size |
|---|---|---|
| Square solve | dimension ≥ 2048 | 512 |
| Lower Cholesky | dimension ≥ 8192 | 2048 |
| Reduced QR | matrix contains ≥ 1048576 elements | 128 |
These defaults can be changed through CNPreferences. The loaded library must support cuSolverMp and Legate must have more than one active GPU. The library selects the algorithm. Legate handles placement, communication, and redistribution. Configure resources before starting Julia, as described in Hardware Configuration.
Ordinary solve and QR tasks remain the fallback. Cholesky also uses the tiled POTRF/TRSM/SYRK/GEMM algorithm for eligible multi-processor configurations. Below its partitioning cutoff this is a single tile. Batched operations still distribute independent matrices, each of which must fit on one processor. SVD and general eigendecomposition do not support distributed factorizations.
Results retain their existing Julia types, element promotion, and shapes. Inputs are preserved. Kernel failures propagate when the runtime reports them. Failed collectives are not retried using another algorithm.
Tuning
Use CNPreferences.set_linalg! to set any of the seven documented constants. MIN_*_MATRIX_SIZE controls when an operation can select cuSolverMp. Solve and Cholesky use the row count. QR uses the number of matrix elements. MIN_SOLVE_TILE_SIZE, MIN_CHOLESKY_TILE_SIZE, and QR_TILE_SIZE set the solver block sizes. QR uses the same block size on both axes.
For the tiled Cholesky fallback, MIN_CHOLESKY_MATRIX_SIZE also sets the single-tile cutoff. MIN_CHOLESKY_TILE_SIZE guides tile subdivision, and MAX_CHOLESKY_TILES_PER_PROC limits the number of tiles per matrix axis relative to the processor count. These are tuning heuristics, not memory limits.
Restart Julia after changing these preferences. The library capabilities and configured resources are determined when the runtime starts.
using CNPreferences
CNPreferences.set_linalg!(; MIN_SOLVE_MATRIX_SIZE=4096, MIN_SOLVE_TILE_SIZE=512)
# Restart Julia to use the new settings.