It becomes a news item: “Linux is twice as fast as any BSD!” Maybe that’s because someone was using a benchmark with a math library that runs poorly on anything that’s not Linux – possibly even through a conscious decision by the programmer to do things a certain way, without thinking of benchmarks, and without the person running the benchmark finding out why.

That’s not evidence, that’s propaganda, and that’s one of the ways it can happen. Yep, I’m also describing a potential situation, not the benchmark from the original link – but one I’ve seen happen.

Also: both machines should be using the same file system and compiler and the programs used were compiled with the same options and are at the same release level and haven’t been tweaked to support specific operating system features of one system and there are no support files present on one system and not the other, and those support libraries are also at the same version, compiled with the same options. It should be pretty clear at this point why accurate benchmarks are difficult.