Benchmarks

Benchmark run collected on 2025-10-19 against the Windows PowerShell host. Raw artifacts live beside this page (bench.json, bench.csv).

Python: 3.13.5 (...\.venv\Scripts\python.exe)
Platform: Windows 11 (build 26200), AMD64, Intel64 Family 6 Model 60 Stepping 3
Timeout: 30 s default
Command under test: $x=1+1; $x | Out-Null

Configuration Highlights

Parameter	Value
Batch sizes	50, 100, 200
Batch repeats	3
Async fan-out (`async_n`)	50
Commands per shell (sequential)	50
Warm shells	0 (single shell reused)
Output files	`bench.json`, `bench.csv`

All timings are wall-clock; throughput reported as completed commands per second.

Size	Sequential mean (ms)	Sequential thrpt	Batch per-cmd mean (ms)	Batch thrpt	Async latency mean (ms)	Async thrpt	Batch eff.	Async eff.
50	4.041	247.30	3.387	295.24	2.359	293.61	1.193	1.186
100	3.754	266.17	3.685	271.40	2.247	304.26	1.019	1.142
200	3.476	287.45	3.203	312.23	1.938	348.66	1.085	1.212

Key takeaways

Async submission sustained up to 349 cmds/s, with sub-2 ms average latency at size 200.
Batch execution remains ~8-19% more efficient than sequential issuance depending on size.
Single-command overhead holds steady around 3.0 ms, implying platform start-up dominates short runs.

Representative workloads (10-20 iterations per type) confirm consistent latency across scenarios.

Command	Mean (ms)	95th percentile (ms)	Est. throughput
Simple echo	3.16 - 10.75	3.54 - 77.07	93 - 316 cmds/s
Math operation	3.47 - 3.70	4.09 - 4.34	271 - 288 cmds/s
Get date	3.70 - 5.78	4.45 - 22.54	173 - 270 cmds/s
File operation	9.08 - 15.47	10.99 - 69.53	65 - 110 cmds/s
Variable assignment	3.38 - 4.22	3.79 - 7.72	237 - 296 cmds/s

Ranges express the min/max across the tested batch sizes.

Session save operations average 302 ms (p50 272 ms, p95 594 ms). Budget roughly 0.3 s when snapshotting state between runs.

Startup cost: 0.33 s for the initial PowerShell host.
Command submission overhead: 3.0 ms independent of payload.
Size scaling: throughput stabilises beyond 50 commands; per-command overhead flattens near 3.2 ms even at 200 commands.

bench.json contains full structured results (env, per-size stats, persistence metrics).
bench.csv offers concise tabular data suitable for spreadsheets.
vs_bench.py is the benchmark harness script.