Skip to content

Latest commit

 

History

History

README.md

Benchmark artifacts

Three independent benchmark campaigns are preserved here. Each directory is self-contained with its own methodology + provenance. Numbers from different campaigns are not blended, every speedup is reported with its source campaign's methodology + caveats.

Directory When Scope Read it for
2026-05_alpha-dev-runs/ May 13-22, multi-date Early integration runs across the ds4 bench harness during alpha development Historical context. The 73.1× canonical headline + 58.7× multi-conversation are both here.
2026-05-24_deployment-mode-matrix/ 2026-05-24, single day 6 deployment modes × 3 restart policies on the same prompt + partial-prefix sweep + scenarios Today's canonical headline matrix. Independent harness, n=5 warmup-primed.
2026-05-24_public-corpus-replay/ 2026-05-24, same day evening extension ShareGPT round-robin + Gutenberg multi-round (real public corpus replay), per-restore + per-save stage timings, raw TCP transport bench Multi-turn realism + diagnostic stage timings. n=1 trial; flag as thin.

How to read each campaign

Each dir contains:

  • README.md, campaign-specific methodology, hardware, software versions, caveats.
  • *.csv / *.jsonl, clean rows, one row per measured cell.
  • raw/ (where applicable), the unmodified bench-harness output JSON.

Charts

assets/ contains one PNG sub-directory per campaign with the per-campaign charts. The chart-generator (scripts/charts/generate_charts.py) reads the CSV/JSONL files in each artifact dir and writes the PNGs back to the matching assets/ subdir.

Why three campaigns instead of one consolidated

Different campaigns answer different questions:

  • Alpha dev runs prove the system worked end-to-end during development. They span ds4 bench harness, multiple dates, and the engineering iteration that produced the integration.
  • Deployment-mode-matrix is the methodical apples-to-apples comparison: same prompt, same machine, same harness, sweep across modes and restart policies. This is the campaign to cite for "vs native ds4" speedup claims.
  • Public-corpus-replay tests realism: real ShareGPT conversations, real Gutenberg long-document QA. This is where exact-restart wins meet messy real workloads, and where save-path overhead surfaces as the daemon tax.

Combining numbers across campaigns would mix prompts, harnesses, and dates. Each campaign is a self-consistent unit; cross-campaign comparisons live in ../BENCHMARKS.md.