benchmarks

UFFS Benchmarks

Benchmark-driven NTFS search. Every number on this page is backed by a raw log, a reproducible script, and a dated frozen report. Nothing here is evergreen — every claim has a version and a date attached to it.

---

Current canonical report

2026-06 · v0.5.120 vs Everything and the UFFS C++ reference →

Four numbers the report establishes on a Ryzen 9 3900XT (cross-tool: 12.8 M records on C/D/F/G; full-scan: all 7 volumes, 25.9 M records):

1. 30 / 30 head-to-head cells faster than Everything at p50 across six pattern classes (exact, prefix, rare-ext, common-ext, regex-alternation, substring) on four drives plus the combined index. Median ratio 0.36× — UFFS is ~2.8× faster. The historical C: prefix statistical tie is gone (80 ms vs 102 ms, 0.78×).
2. Every cell published in the previous (v0.5.66) snapshot got faster — median −33% — while Everything's own numbers held roughly flat. The gap widened, not narrowed.
3. Full-scan export is a workload Everything cannot run (es.exe ~2 GB IPC export ceiling): UFFS streams the complete 23.3 M-row estate (all 7 volumes) to CSV in 12.0 s ≈ 1.95 M records/sec — the April snapshot scale, 12% faster, +13% throughput.
4. 180×–3 400× faster than the C++ reference on targeted queries (daemon HOT vs per-invocation MFT re-read); 6.6× on combined full-scan — and the combined-drive regex cell DNF'd the C++ tool entirely (> 120 s vs UFFS 43 ms).

The report publishes everything these numbers don't cover too — the zero-match G-drive caveat, the C++ row-count divergences, and what this benchmark explicitly does not claim. The two v0.5.66-era known regressions (* top-100 and --sort path vs the v0.5.4 baseline) remain tracked in the archived April report; they were not re-measured in this snapshot.

---

How UFFS benchmarks

Ready to run a benchmark cycle? See the operator runbook → for prerequisites, step-by-step commands, crash recovery, and how to promote results to a canonical report.

Four fairness principles, documented in full in methodology.md (the single-link reply to "this comparison is rigged because..."):

• Separate cold / warm / hot. Cold build + warm restart + hot query are three different workloads. We measure and publish them separately instead of averaging them into one "startup time" lie.
• Separate interactive from bulk. Targeted-query latency (notepad.exe, *.dll) and full-scan export (* → CSV for 23 M rows) are different workload classes. Different tools win each. We test both.
• Publish the failures. When a workload regresses against our own prior baseline it gets named, measured, root-caused, and tracked (see §Known regressions in the archived 2026-04 report for the two v0.5.66-era examples).
• Publish the raw data. Every table above and in the canonical report cites the exact log file and line range. The curated, verbatim raw captures live in raw/ (git-tracked, never edited after commit); all benchmark scripts under scripts/windows/. Click any citation in the canonical report to land on the actual PowerShell log line that produced the number.

---

Reproduce

Elevated PowerShell, repository root, after cargo build --release:

# The full benchmark suite: drive negotiation (Everything RAM budget), the
# cross-tool harness (UFFS Rust vs UFFS C++ vs Everything), parity, native
# full-suite, REPORT-DRAFT.md assembly, and the brand-kit charts.
just bench-suite --drives C,D,F,G

The suite writes a dated bundle containing REPORT-DRAFT.md, cross-tool-summary.csv, the three competition charts, and a ## vs baseline section comparing the run against this page's current canonical numbers (baseline.json). Promotion into a canonical report is the reviewed copy-edit of that draft. Individual harnesses remain runnable standalone — see scripts/windows/cross-tool-benchmark.rs.

---

Archive

Frozen snapshots of prior canonical reports, never retroactively edited. See archive/README.md for the archive policy.

• 2026-04 · v0.5.66 vs Everything and the UFFS C++ reference — the first canonical snapshot (12/12 cells vs Everything on C+D, median 0.51×; cold-parity and memory-scaling sections; the two tracked regressions). Superseded by the v0.5.120 report above.

---

What's not in here

• No "fastest on Windows" superlative. Different workloads have different winners. UFFS is measurably faster than Everything and the C++ reference on the documented patterns — that's what we claim.
• No WizFile numbers yet — WizFile has no automated-comparison flag (--out equivalent). When we add harness support, results land here.
• No "bring your own drives" comparison — all numbers come from the same test machine. Per-user results will vary with CPU, drive topology, and drive fullness.

For the full positioning story (competitor landscape, claim framework, what to amplify vs qualify), see the internal strategy docs: docs/dev/architecture/ntfs_mcp_marketing_strategy_deep_dive.md and docs/dev/architecture/marketing_strategy_adjustment_after_benchmark_update.md.