Skip to content

Add benchmark tables comparing e57 with pye57 to the README - #53

Merged
dancergraham merged 1 commit into
dancergraham:mainfrom
mohsaad:docs/benchmark-tables
Nov 16, 2026
Merged

dancergraham merged 1 commit into
dancergraham:mainfrom
mohsaad:docs/benchmark-tables

Conversation

@mohsaad

@mohsaad mohsaad commented Nov 11, 2026

Copy link
Copy Markdown
Contributor

Relates to #21 (RAM usage).

Adds a Performance section to the README. It compares e57 0.2.3 with pye57 0.4.19 on read time, peak memory, output size and overhead (peak memory divided by output size).

What's measured

  • Files:
    • a real scan: gargage.e57 from the libE57 example data, 1 scan, 27.8M points
    • two synthetic files written with pye57: 50M points in 4 scans, and 1B points in 20 scans
  • Machines: an Apple Silicon Mac (Docker, linux/arm64) and an AWS r7i.8xlarge (amd64, 256 GiB).
  • Output: every reader produces what read_points returns: all scans merged, pose applied, invalid points dropped.
  • Two pye57 rows: pye57 returns one scan at a time, so it's shown with a naive np.concatenate merge and with a preallocated merge. The merge strategy changes its peak memory a lot: 60.5 GiB naive vs 33.0 GiB preallocated at 1B points.
  • Isolation: each read runs in its own process after a warm-up read, so the file is already in the page cache. Times are the median of 3 runs. The correctness check runs in a separate process too, so it can't inflate the peak-memory numbers.
  • Correctness: both libraries returned the same data in every run.

Summary

  • e57 adds almost no memory on top of its output (1.00–1.02×). On a file with one big scan it uses about 2.6× less memory than pye57.
  • With a preallocated merge, pye57's overhead shrinks on multi-scan files. At 1B points it peaks below e57, mostly because e57 returns color as float32 (12 bytes per point vs 3).
  • pye57 reads 20–25% faster on Intel.

The README presents these results as they are, including where pye57 does better. Converting color to uint8 or making the color type optional would close most of the memory gap at large sizes. That could be a follow-up if you're interested.

Notes

🤖 Generated with Claude Code

@dancergraham dancergraham left a comment

Copy link
Copy Markdown
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Love this thank you !

@mohsaad
mohsaad marked this pull request as ready for review November 15, 2026 18:06
Adds a Performance section to the README with read time, peak memory,
output size and overhead for e57 0.2.3 and pye57 0.4.19. It covers a real
27.8M-point scan and synthetic files of 50M and 1B points, on Apple
Silicon and on an AWS r7i.8xlarge.

pye57 is shown with both a naive np.concatenate merge and a preallocated
merge, since its peak memory depends heavily on how scans are combined.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@mohsaad
mohsaad force-pushed the docs/benchmark-tables branch from a65cce9 to 03268c1 Compare November 15, 2026 21:08
@dancergraham
dancergraham merged commit 343d51e into dancergraham:main Nov 16, 2026
17 checks passed
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.

2 participants