Skip to content

skill(faer-optimization): parallel splits must scale past the development host #96

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
shinaoka merged 1 commit into from
Oct 28, 2026
Merged

skill(faer-optimization): parallel splits must scale past the development host #96

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 16 additions & 0 deletions skills/faer-optimization/SKILL.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,6 +46,22 @@ reached through `skills/faer-optimization/SKILL.md` and the per-harness links be
with the number that rejected them, verification, campaign), the `PATCHSET.md` row, and a
lesson that will outlive the change into `PERFORMANCE_TIPS.md`.

## Parallel splits must scale

Before measuring any new or changed parallel region, check it against the *Parallel Scaling* list
in `rules/common/performance.md` (tensor4all-agent-rules) at the largest thread count faer is run
with, 64 and more, not only at evo's 8. In this fork the recurring traps are:

- a column split where every task reads the whole Householder basis or panel (`p × size` traffic,
tensor4all/faer#90); pack or stage the shared operand once, or split in two dimensions;
- a row split into one partial per thread with a serial combine (`O(p)` memory and combine);
- a constant task cap ("at most four tasks") or a threshold tuned at one thread count.

Name the cost of the split as a function of `p` in the worklog. Scaling evidence goes beyond evo:
run 1, 8 and the largest count on a many-core host (broccoli, 64 cores) and compare with the
OpenBLAS arm. A split that only wins at 8 threads, or only against a starved baseline, is not
adopted.

## The yardstick is not a source

The vendor is timed through its public API only. No disassembly, no profiler into its internals,
Expand Down
Loading