Guidelines

This page collects practical setup rules that are repeatedly referenced in tutorials.

Parallelization and Block Size

• Keep block_size <= 16 in most runs. Larger blocks can increase mapping parallelism, but factorization and diagonalization costs can grow too much.
• A useful scaling regime is:

\[\frac{\mathrm{howmany}}{\mathrm{block\_size}} \gtrsim 100.\]

In this regime, column-level mapping parallelization is usually efficient.

Matrix Size Heuristic

• Around matrix size ~250_000, MulColsParallel is often the best default strategy on CPU.
• For larger matrices, memory pressure and runtime increase significantly. This is where TwoLevelParallel often becomes useful, especially with Block Lanczos.

howmany vs Memory Tradeoff

Increasing howmany can improve robustness and reduce restart pressure, but it also raises memory usage and can make projected-eigensolver steps more expensive.

Use display_report to monitor matrix multiplications, factorization progress, timing split, and order_safety_factor while tuning.

This page was generated using Literate.jl.