It hardly seems strange if Slava's main computer happens to be a mac.
Do you have some reason to expect that the results would be substantially different on some other platform? As noted in the article these are mostly srather narrow pure computation benchmarks, so the OS shouldn't matter a lot. Sure, the OS X kernel is crap when it comes to supporting exotic features, so SBCL has to do a little bit of extra work around bugs in things like signal handling. And I could believe using mprotect() for a hardware write barrier have a worse relative performance compared to a software barrier on OS X than on Linux. But at least in the context of pure computation these kinds of effects shouldn't distort things all that much.
