The test does not look realistic: https://github.com/Cranot/grouped-simd-hashtable/blob/master...

Better to use a few distributions of keys from production-like datasets, e.g., from ClickBench. Most of them will be Zipfian and also have different temporal locality.