[gspark-etri]

Background

onggi's numbers for the ATOM Llama-3.1-8B build set had no baselines on the same data (#51).

Changes

• bench/baselines.sh: on one build directory it measures storage, the transfer for a variant switch, and times:
  • raw;
  • OCI-style layers;
  • one tar with zstd --long=31;
  • fixed-block dedup (64 KiB, 1 MiB);
  • FastCDC 64 KiB dedup;
  • Xet-style CDC with per-chunk compression (onggi --sub 0);
  • onggi.
• docs/evaluation.md: the results.
• README in both languages: status P2 (the CSI driver, node cache, peers and delta apply exist) and a link to the evaluation.

Verification

• A small synthetic run on the node gave sane rows. The full run on the 263.9 GB ATOM set:

method	stored	vs raw
OCI layers, zstd -3	199.2 GB	1.33×
tar, zstd -3 --long=31	199.1 GB	1.33×
fixed 1 MiB / 64 KiB dedup	116.4 / 113.4 GB	2.27× / 2.33×
FastCDC 64 KiB dedup	46.8 GB	5.64×
Xet-style CDC + per-chunk compression	29.5 GB	8.95×
onggi	17.0 GB	15.6×

decode B2 → B8 on a node that holds B2	transferred
OCI layer	11.46 GB
Xet-style	6.1 MB
onggi	28.3 MB

onggi sends more than Xet-style here because it compresses ~4 KiB sub-chunks one by one. #52 tracks that.

Compatibility and limits

• Docs and a script only. Xet itself was not run (it needs the HF service); the Xet-style row is an approximation and is labeled as such.

Related Issue

Closes #51

🤖 Generated with Claude Code