Skip to content

perf: measured NumPy table at v0.41.0 on one Zen 3 core - #119

Merged
tannevaled merged 2 commits into
mainfrom
perf/numpy-table-v0.41.0
Nov 14, 2026
Merged

tannevaled merged 2 commits into
mainfrom
perf/numpy-table-v0.41.0

Conversation

@tannevaled

Copy link
Copy Markdown
Contributor

This is the first measured docs/perf.json / docs/perf-table.md. The table had been "deferred" since the harness landed (Refs #105), because the development host's load was above the gate of 8. The figures are on #118.

  • Commit measured: v0.41.0, de3f41ac4ec1e21554a63f95e4041143acfd58bd, clean. The table's Commit line shows no -dirty.
  • Host: cfarm420 (AMD EPYC 7773X, Zen 3), taskset -c 9, GOMAXPROCS=1, Go 1.27.4.
  • Reference: NumPy 2.5.4, SciPy 1.18.1, scipy-openblas 0.3.34 at one thread, Python 3.12.12, all through uv from bench/refs.txt.
  • Load over the rounds: 4.12 / 4.48 / 5.56 (min / median / max). All 256 rows are valid, with 0 retries.
  • How it was run:
    • The test binary was cross-compiled from a clean v0.41.0 worktree.
    • The tag reached the host as a git bundle, because cfarm420's HTTPS to github.com was reset.
    • The binary ran from outside the clone, so nothing untracked made the tree dirty.
    • An earlier attempt with the binary inside bench/ gave the same verdicts, stamped …-dirty, and was discarded.
  • Result: 188 slower, 54 faster, the rest at parity. perf: 188 of 256 rows slower than NumPy 2.5.4 on one Zen 3 core (v0.41.0), by cause #118 tracks the slower rows by cause.
  • docs/perf.md: the Status section now describes the measured table, and the harness is named refbench v0.3.0.

A refbench behaviour hit on the way, to report to go-scikit/refbench:

  • A partial -perf.run (load gate tripped) writes docs/perf.json into the clone.
  • -perf.resume then sees the tree as dirty.
  • -perf.commit <tag> refuses a dirty tree.

A partial run therefore cannot be resumed, and a fresh -perf.run -perf.onload=wait was the way through.

Refs #118

🤖 Generated with Claude Code

tannevaled and others added 2 commits November 14, 2026 22:10
…54 faster; Refs #118)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…darray: performance is open (#118)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@tannevaled
tannevaled merged commit 91bdd42 into main Nov 14, 2026
13 checks passed
@tannevaled
tannevaled deleted the perf/numpy-table-v0.41.0 branch November 14, 2026 21:21
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant