Open-addressing hash map

Hand-rolled FNV-1a hash table, 100k mixed put/get/delete operations. No built-in map allowed.

Runtime Compile time

Runtime · median per inner-loop window

median of 10 runs

Native Gocompiled

22.5 ms0.04×

Piko interpbytecode VM

506 msbaseline

CPython 3.13bytecode VM

673 ms1.33×

PyPy 7.3tracing JIT

89.1 ms0.18×

Ttengobytecode VM

1.71 s3.39×

Sscriggobytecode VM

unsupported in scriggo

n/a

Mmvmbytecode VM

1.05 s2.08×

YyaegiAST walker

1.76 s3.49×

Full statistics

Runner	N	Compile	Runtime	P95	Stddev	RSS	vs piko	Status
Native Gocompiled	10	181 ms	22.5 ms	22.7 ms	191 µs	68 MiB	0.04×	OK
Piko interpbytecode VM	10	2.58 ms	506 ms	508 ms	1.04 ms	313 MiB	1.00×	OK
CPython 3.13bytecode VM	10	627 µs	673 ms	684 ms	9.35 ms	n/a	1.33×	OK
PyPy 7.3tracing JIT	10	496 µs	89.1 ms	93.1 ms	1.45 ms	n/a	0.18×	OK
tengobytecode VM	10	372 µs	1.71 s	1.78 s	34.1 ms	3.07 GiB	3.39×	OK
scriggobytecode VM	0	n/a	n/a	n/a	n/a	n/a	n/a	unsupported
mvmbytecode VM	10	679 µs	1.05 s	1.08 s	11.0 ms	70 MiB	2.08×	OK
yaegiAST walker	10	812 µs	1.76 s	1.90 s	103 ms	74 MiB	3.49×	OK

Workload & symmetry rules

Workload

Build a HashTable with parallel keys[], values[], and state[] arrays (0=empty, 1=occupied, 2=tombstone). Linear probing on collision, doubles capacity when load > 70%. Hash function is byte-by-byte FNV-1a. A deterministic LCG drives 100,000 mixed operations.

Symmetry rules

No built-in map / dict anywhere in the hot path.
No built-in or stdlib hash functions; FNV-1a must be hand-rolled.
Power-of-two capacity, linear probing, tombstone deletes: every runner uses the same algorithm.

Why this benchmark exists

Pure interpreter throughput without the cheat of a C-level hash table. Reveals what the interpreter's array / arithmetic / loop performance looks like under sustained load.

Source code

piko / Go
piko_source.go
native Go
native_main.go
CPython / PyPy
cpython.py
tengo
script.tengo