Parallel word count (Monte Cristo)

Parallel word-frequency count over ~2.7 MB of Project Gutenberg text. Each runner uses its idiomatic CPU-parallel primitive.

Runtime · median per inner-loop window

median of 10 runs

Native Gocompiled

19.0 ms0.18×

Piko interpbytecode VM

108 msbaseline

CPython 3.13bytecode VM

181 ms1.68×

PyPy 7.3tracing JIT

353 ms3.28×

Ttengobytecode VM

unsupported in tengo

n/a

Sscriggobytecode VM

unsupported in scriggo

n/a

Mmvmbytecode VM

302 ms2.80×

YyaegiAST walker

unsupported in yaegi

n/a

Runner	N	Compile	Runtime	P95	Stddev	RSS	vs piko	Status
Native Gocompiled	10	181 ms	19.0 ms	21.2 ms	1.35 ms	72 MiB	0.18×	OK
Piko interpbytecode VM	10	4.11 ms	108 ms	110 ms	2.48 ms	418 MiB	1.00×	OK
CPython 3.13bytecode VM	10	671 µs	181 ms	184 ms	1.70 ms	n/a	1.68×	OK
PyPy 7.3tracing JIT	10	540 µs	353 ms	359 ms	5.10 ms	n/a	3.28×	OK
tengobytecode VM	0	n/a	n/a	n/a	n/a	n/a	n/a	unsupported
scriggobytecode VM	0	n/a	n/a	n/a	n/a	n/a	n/a	unsupported
mvmbytecode VM	10	679 µs	302 ms	307 ms	4.57 ms	93 MiB	2.80×	OK
yaegiAST walker	0	n/a	n/a	n/a	n/a	n/a	n/a	unsupported

Workload & symmetry rules

Workload

Load The Count of Monte Cristo (Project Gutenberg #1184) corpus from disk.
Split into 16 word-aligned chunks.
Spawn 16 workers (goroutine + WaitGroup on Go/piko; multiprocessing.Pool on Python).
Each worker tokenises via byte-walk and builds a local word → count map.
Merge maps, compute top-50 by (count desc, word asc) via hand-rolled insertion-sort.

Each runner uses its idiomatic CPU-parallel primitive. Python's multiprocessing.Pool (not threading, because of the GIL).
No str.split, no regexp, no sort.*/sorted(key=...): tokenisation and ranking are hand-rolled.

Directly measures the GIL-free advantage. Piko inherits Go's goroutine scheduler; CPython has to fork.

Source code