Skip to content

Benchmarks ​

matchmaker is built for rich multi-column, ANSI-styled, and wrapped terminal UIs without giving up interactive speed. ResultsUI double-buffers rendered table cells and column widths (row_cache) and skips table rebuilds when no state is invalidated (is_clean), while supporting both the nucleo and SIMD-accelerated frizbee matcher backends.

Methodology & environment ​

  • Machine: AMD Ryzen 9 7950X (16C/32T), Linux 7.3.4
  • Binaries: mm 0.1.7 built with --profile dist (codegen-units = 1, lto = "fat"), fzf 0.75.0, sk 0.20.5
  • PTY harness: benches/bench_tui.py in a 120x40 PTY (start.sync = true / --sync, tui.sleep_ms = 0 on mm, FZF_DEFAULT_OPTS="" on fzf)
  • Aggregation: median of 5 measured runs after 1 warmup run

Heavy 6-column ANSI + wrapped viewport (200k rows) ​

Each row contains ANSI color escapes and 6 tab-separated columns. mm runs with ansi = true, results.wrap = true, and max_height = 4 (full column auto-sizing and line wrapping), while the other pickers run with --ansi and single-line truncation.

PickerFirst frameBroad query50× cursor navSwitch query (Ctrl-U)Accept (Enter)Peak RSS
mm (nucleo)240.9 ms3.7 ms5.1 ms4.6 ms20.1 ms124.3 MB
mm (frizbee)218.0 ms5.3 ms4.5 ms4.5 ms18.4 ms123.1 MB
fzf --ansi622.3 ms5.4 ms4.6 ms2.3 ms6.0 ms117.1 MB
sk --ansi1103.9 ms13.0 ms381.4 ms17.0 ms10.9 ms175.9 MB
  • 2.6–2.9× faster to first frame than fzf and 4.6–5.1× faster than sk on 200k ANSI-colored multi-column rows, even with wrapping and column layout enabled.
  • Cached cursor navigation: 50 cursor moves (25× Down + 25× Up) complete in 4.5–5.1 ms on mm (vs 381.4 ms on sk).

500k scale & query-shape transitions ​

500,000 file paths with synchronous ingestion (start.sync = true / --sync), measuring first frame, incremental typing (g → wp → robe), switching to a disjoint unique match (Ctrl-U + zxq_validator_9), switching back to the 499-hit query (Ctrl-U + gwprobe), and peak RSS.

PickerFirst frameIncremental (gwprobe)Switch (zxq_validator_9)Broad re-filter (gwprobe)Peak RSS
mm (frizbee)88.8 ms22.4 ms20.2 ms24.5 ms109.9 MB
mm (nucleo)94.7 ms20.6 ms19.2 ms34.7 ms109.5 MB
fzf530.0 ms16.5 ms4.4 ms1.3 ms99.2 MB
sk645.0 ms32.5 ms22.7 ms18.5 ms225.6 MB
  • On 500k paths, mm renders its first frame 5.6–6.0× faster than fzf and 6.8–7.3× faster than sk, while keeping peak RSS (109.5–109.9 MB) less than half of sk (225.6 MB).

Synchronous select-1 / exit.first startup (200k paths) ​

Startup, full ingestion (start.sync = true / --sync), matching a unique initial query across 200,000 plain paths, and exiting automatically (exit.first = true / -1).

PickerWall timeCPU timePeak RSS
fzf --sync -120.9 ms32.2 ms45.3 MB
mm (frizbee)37.5 ms154.9 ms94.2 MB
mm (nucleo)37.8 ms134.0 ms75.9 MB
sk --sync -1219.3 ms419.2 ms75.6 MB

matchmaker-lib microbenchmarks ​

Measured directly in Rust without PTY overhead via cargo bench -p matchmaker-lib --bench ui_and_matcher (matchmaker-lib/benches/ui_and_matcher.rs).

ResultsUI::update_table caching (120×40, 6-col ANSI, wrap = true) ​

PathMedianMinp95Speedup vs cold
cold_miss (set_dirty)339.1 µs332.6 µs353.1 µs1.0×
row_cache (cursor navigation)16.3 µs14.8 µs19.2 µs20.8×
is_clean (no-op tick)< 0.1 µs< 0.1 µs< 0.1 µs—

nucleo vs frizbee query throughput & typo tolerance ​

DatasetBackendIngest1-char (g)3-char (gwp)7-char (gwprobe)15-charBackspace (gwp)1-typo (zxq_validatxr_9)
200knucleo29.6 ms4.18 ms7.36 ms7.27 ms6.42 ms7.22 ms6.21 ms (0 hits)
200kfrizbee28.3 ms7.15 ms7.20 ms7.40 ms6.55 ms7.37 ms6.29 ms (0 hits)
200kfrizbee (max_typos=1)27.7 ms14.20 ms7.18 ms7.33 ms6.92 ms7.07 ms7.22 ms (1 hit)
500knucleo72.4 ms16.78 ms16.84 ms16.62 ms15.16 ms16.19 ms15.05 ms (0 hits)
500kfrizbee71.6 ms16.34 ms16.22 ms16.14 ms15.32 ms16.92 ms15.72 ms (0 hits)
500kfrizbee (max_typos=1)70.2 ms44.60 ms16.71 ms16.61 ms15.06 ms16.58 ms15.62 ms (1 hit)

Reproducing locally ​

sh
cargo build --profile dist -p matchmaker-cli
python3 benches/bench_tui.py --mm ./target/dist/mm --items 200000 --scale-items 500000 --runs 5 --warmup 1
cargo bench -p matchmaker-lib --bench ui_and_matcher