Benchmarks

All benchmarks run on an AMD Ryzen 9 9900X 12-Core Processor (24 cores, 121 GB RAM), averaged over 100 runs after 10 warmup iterations. Times in milliseconds (lower is better) unless noted otherwise.

---

Startup Latency

How fast each shell can execute shell -c 'true' and exit.

shell -c 'true' — round-trip

data

dash

0.26ms

bash

0.43ms

sh

0.47ms

lash

0.55ms

zsh

0.56ms

fish

6.5ms

shell -c 'echo x' — round-trip

data

dash

0.21ms

sh

0.42ms

bash

0.43ms

zsh

0.51ms

lash

0.57ms

fish

6.4ms

shell -c 'echo x | cat' — round-trip

data

dash

0.61ms

bash

0.74ms

lash

0.75ms

sh

0.77ms

zsh

0.91ms

fish

6.1ms

---

Pipe Throughput

Raw data throughput through pipes. MB/s charts are higher-is-better.

64 MB single pipe — throughput

data

lash

8878 MB/s

lash-turbo

8745 MB/s

dash

8325 MB/s

zsh

7566 MB/s

bash

7409 MB/s

sh

7073 MB/s

fish

4257 MB/s

1 GB single pipe — throughput

data

lash-turbo

8669 MB/s

bash

7375 MB/s

sh

7312 MB/s

zsh

7309 MB/s

lash

7240 MB/s

dash

7107 MB/s

fish

6955 MB/s

64 MB through 3 cat stages

data

lash-turbo

7.4ms

sh

8.3ms

bash

8.8ms

lash

8.9ms

zsh

8.9ms

dash

9.3ms

fish

16ms

16 MB streamed to sink

data

dash

2.3ms

lash-turbo

2.6ms

bash

2.7ms

sh

2.7ms

zsh

2.9ms

lash

2.9ms

fish

8.9ms

echo | cat — command latency

data

dash

0.46ms

lash-turbo

0.63ms

lash

0.68ms

bash

0.71ms

sh

0.71ms

zsh

1.0ms

fish

6.0ms

Pipe Chain Depth

echo through 5 cat stages

data

dash

0.79ms

lash-turbo

0.91ms

lash

0.93ms

bash

1.1ms

sh

1.1ms

zsh

1.4ms

fish

7.0ms

echo through 10 cat stages

data

dash

0.98ms

lash-turbo

1.1ms

bash

1.2ms

lash

1.3ms

sh

1.4ms

zsh

1.9ms

fish

8.1ms

Pipe Scaling (16 MB)

16 MB — 1 cat stage

data

lash-turbo

2.6ms

dash

2.7ms

sh

2.9ms

bash

3.0ms

zsh

3.0ms

lash

3.1ms

fish

9.5ms

16 MB — 4 cat stages

data

lash-turbo

2.3ms

dash

3.4ms

sh

3.6ms

bash

3.8ms

lash

3.8ms

zsh

4.0ms

fish

11ms

16 MB — 8 cat stages

data

lash-turbo

2.4ms

dash

3.6ms

bash

5.2ms

lash

5.3ms

zsh

5.9ms

sh

6.7ms

fish

11ms

16 MB — 16 cat stages

data

lash-turbo

2.6ms

bash

6.8ms

dash

7.0ms

lash

7.0ms

zsh

8.5ms

sh

8.9ms

fish

14ms

File I/O

16 MB write to file

data

sh

2.5ms

zsh

2.6ms

dash

3.1ms

bash

3.1ms

lash

3.1ms

lash-turbo

3.6ms

fish

9.4ms

16 MB pipe to file

data

lash-turbo

3.9ms

dash

4.5ms

bash

4.5ms

sh

4.5ms

lash

4.6ms

zsh

4.7ms

fish

12ms

16 MB read from file through pipe

data

lash-turbo

6.9ms

sh

7.0ms

dash

7.2ms

lash

7.5ms

zsh

7.5ms

bash

7.6ms

fish

14ms

16 MB to /dev/null (overhead baseline)

data

dash

0.56ms

sh

0.82ms

bash

0.84ms

lash-turbo

0.89ms

zsh

0.90ms

lash

0.93ms

fish

6.7ms

---

Turbo Mode

Turbo mode rewrites common pipelines into native array operations — no fork/exec overhead. Speedups are turbo vs forked lash.

seq 1M | sort | tail — turbo 6.9x vs forked

data

lash-turbo

39ms

lash

298ms

fish

300ms

bash

305ms

zsh

307ms

dash

308ms

sh

314ms

sort 100K lines — turbo 3.9x vs forked

data

lash-turbo

6.9ms

lash

27ms

sh

27ms

zsh

28ms

dash

28ms

bash

29ms

fish

35ms

sort | head from 100K — turbo 7.7x vs forked

data

lash-turbo

4.0ms

zsh

24ms

lash

24ms

sh

24ms

bash

24ms

dash

24ms

fish

30ms

sort | tail from 100K — turbo 5.2x vs forked

data

lash-turbo

4.4ms

lash

26ms

sh

26ms

bash

26ms

dash

26ms

zsh

26ms

fish

32ms

grep | sort | head from 100K — turbo 2.6x vs forked

data

lash-turbo

3.7ms

dash

10.0ms

lash

10ms

sh

10ms

bash

11ms

zsh

11ms

fish

18ms

sort 1M lines reverse numeric — turbo 7.8x vs forked

data

lash-turbo

40ms

bash

305ms

lash

307ms

fish

313ms

sh

315ms

dash

316ms

zsh

320ms

sort+head+sort+tail from 100K — turbo 6.2x vs forked

data

sh

39ms

zsh

39ms

bash

41ms

lash

41ms

dash

41ms

fish

47ms

lash-turbo

494ms

5-stage pipeline on 500K — turbo 4.2x vs forked

data

zsh

101ms

sh

103ms

dash

108ms

lash

109ms

bash

109ms

fish

112ms

lash-turbo

3154ms

generate+sort+uniq+sort from 100K — turbo 6.5x vs forked

data

lash-turbo

36ms

lash

193ms

bash

194ms

sh

197ms

zsh

203ms

dash

203ms

fish

215ms

---

Scripting Operations

Common data-processing patterns across shells.

sort 1K lines — reverse numeric

data

lash-turbo

0.65ms

dash

0.66ms

lash

0.96ms

bash

1.0ms

sh

1.0ms

zsh

1.2ms

fish

6.9ms

sort 10K lines — reverse numeric

data

lash-turbo

1.2ms

dash

2.8ms

lash

3.0ms

sh

3.0ms

bash

3.0ms

zsh

3.3ms

fish

10ms

grep filter 1K lines

data

lash-turbo

0.59ms

dash

0.62ms

bash

0.83ms

lash

0.86ms

sh

0.93ms

zsh

1.3ms

fish

6.9ms

awk filter 10K lines

data

dash

1.8ms

sh

2.0ms

lash

2.0ms

bash

2.0ms

lash-turbo

2.2ms

zsh

2.3ms

fish

7.7ms

awk map (x2) 10K lines

data

dash

2.2ms

sh

2.3ms

bash

2.4ms

lash

2.4ms

lash-turbo

2.5ms

zsh

2.6ms

fish

9.7ms

grep pattern in 10K lines

data

dash

0.64ms

lash-turbo

0.75ms

sh

0.89ms

bash

0.98ms

lash

1.0ms

zsh

1.1ms

fish

6.8ms

100K lines through grep filter

data

dash

2.0ms

lash

2.1ms

bash

2.2ms

sh

2.3ms

zsh

2.5ms

lash-turbo

3.5ms

fish

11ms

100K small lines through pipe

data

dash

1.1ms

lash-turbo

1.2ms

lash

1.4ms

bash

1.5ms

sh

1.8ms

zsh

2.0ms

fish

8.9ms

1M small lines through pipe

data

dash

4.3ms

lash-turbo

4.3ms

sh

4.6ms

zsh

4.8ms

bash

4.8ms

lash

4.9ms

fish

12ms

---

Process Spawn Overhead

How fast each shell can fork and exec processes.

single fork+exec (no pipe)

data

dash

0.23ms

bash

0.47ms

sh

0.51ms

zsh

0.64ms

lash-turbo

0.64ms

lash

0.72ms

fish

6.7ms

2-stage no-op pipe setup

data

dash

0.41ms

sh

0.67ms

bash

0.68ms

lash-turbo

0.71ms

lash

0.74ms

zsh

0.77ms

fish

6.9ms

5-stage no-op pipe setup

data

dash

0.54ms

zsh

0.82ms

sh

0.84ms

bash

0.85ms

lash-turbo

0.91ms

lash

0.98ms

fish

6.8ms

10-stage no-op pipe setup

data

dash

0.69ms

bash

1.0ms

sh

1.1ms

lash-turbo

1.1ms

lash

1.1ms

zsh

1.5ms

fish

6.4ms

---

How Turbo Mode Works

Turbo mode applies these optimizations automatically:

• Passthrough stripping — removes identity operations so they never execute
• Numeric sort key pre-computation — pre-computes keys in O(N) instead of parsing inside the comparator at O(N log N)
• Streaming wc -l — counts newlines in the byte stream without collecting lines
• C strtod for numeric conversion — calls C’s strtod directly, avoiding D’s to!double exception overhead
• Fused operations — grep | head, grep | tail, and grep | wc run in a single pass over the data

Running Benchmarks

dub run :benchmarks

Options

Flag	Description
--runs N	Number of iterations per scenario
--warmup N	Warmup iterations before measurement
--scenario S	Run only the named scenario
--json	Output results in JSON format
--verbose	Print per-iteration timings

To reproduce these numbers:

dub run :benchmarks -- --runs 100 --warmup 10 --verbose