Skip to content

benchmarks.md

Latest commit

 

History

History
111 lines (75 loc) · 4.08 KB

benchmarks.md

File metadata and controls

111 lines (75 loc) · 4.08 KB
Raw

Benchmark Results

Benchmarks run on macOS with ~90 git repositories, using hyperfine with minimum 3-5 runs per configuration.

Environment

  • macOS (Darwin 24.6.0)
  • 20 CPU cores
  • SSH multiplexing enabled (ControlMaster)
  • Rust implementation (rust)

Baseline

script/bench compare also benchmarks bin/git-all-gnu-parallel, a shell baseline that discovers the same depth-1 repos and runs one git -C <repo> command per repository through GNU Parallel. It is intended as a simple comparison point for orchestration overhead, not as a feature-complete implementation.

git status

Local filesystem operation - bottleneck is process spawning and filesystem I/O.

Workers Time Notes
1 2.18s
2 1.27s
4 0.85s Optimal
8 0.98s Contention starts
12 1.17s
16 1.80s
20 (auto) 1.62s

Finding: 4 workers is optimal for local operations. More parallelism causes filesystem/process contention.

git pull

Network-bound operation - bottleneck is SSH connection and remote server.

Workers Time Notes
1 26.5s
2 15.1s
4 8.4s
8 5.5s Optimal

Finding: 8 workers is optimal for network operations. More parallelism hides network latency.

git fetch

Similar to pull - network-bound.

Workers Time Notes
1 25.0s
4 8.2s
8 5.3s Optimal

SSH Multiplexing Impact

Without SSH multiplexing, git pull with 4 workers took ~25s. With multiplexing enabled, it dropped to ~8s - a 3x improvement.

See README.md for SSH multiplexing configuration.

Recommendation

Default worker count: 8

This provides the best balance:

  • Near-optimal for network operations (pull, fetch)
  • Acceptable for local operations (status is ~1s vs 0.85s optimal)

TTY Benchmarking

The crossterm renderer is only exercised when stdout is a TTY. Use script/bench git --tty to wrap benchmarked commands in a pseudo-terminal:

script/bench git -I rust -b main -t crossterm-smart-tty -d ~/work -c status -n 8 --tty
script/bench git -I rust -b main -t crossterm-smart-tty -d ~/work -c pull -n 8 --tty
script/bench git -I rust -b main -t crossterm-smart-tty -d ~/src/oss -c status -n 8 --tty
script/bench git -I rust -b main -t crossterm-smart-tty -d ~/src/oss -c pull -n 8 --tty

Run the same matrix without --tty to verify non-TTY behavior remains stable.

Crossterm TTY Table Validation

Validation run on April 23, 2026 with script/bench git -I rust -b main -t HEAD -n 8. Benchmarks used -w 1 -m 3 unless noted otherwise.

Non-TTY

  • ~/work status: main 3.157s vs current 3.169s, effectively neutral.
  • ~/work pull: rerun with -w 2 -m 5 due to an earlier noisy sample; main 7.272s vs current 7.478s, current about 3% slower.
  • ~/src/oss status: main 4.392s vs current 4.334s, current about 1% faster.
  • ~/src/oss pull: main 2.851s vs current 3.034s, current about 6% slower.

TTY

  • ~/work status: main 3.459s vs current 3.453s, effectively neutral.
  • ~/work pull: main 7.366s vs current 7.650s, current about 4% slower.
  • ~/src/oss status: main 4.688s vs current 4.694s, effectively neutral.
  • ~/src/oss pull: main 3.446s vs current 3.190s, current about 7% faster, though the baseline side had high variance.

Visual Checks

  • ~/work status: tmux live/final capture reviewed. Live pane showed running rows with no header row, and the final table remained on screen.
  • ~/src/oss status: tmux live/final capture reviewed. Live pane showed running rows with no header row, and the final table remained on screen.
  • ~/work pull: tmux live/final capture reviewed. Final table kept inline pull errors in the output column and remained on screen.
  • ~/src/oss pull: tmux live/final capture reviewed. Final table kept inline tracking-info errors in the output column and remained on screen.

Trace Spot Check

  • status ~/work: repos=100 first_exit_ms=158 first_print_ms=515 delayed_repos=94 max_ordered_wait_ms=1534 total_ms=3447