If the benchmark is bound by the database speed, wouldn't the expected result be that all implementations returned roughly the same number of requests per second?