...in long-running loops. I wonder how large the start-up costs are, though? Granted, non-long-running loops are by definition not running for long, so it's ok if they're a bit less efficient, but 2x faster in hotpots is still different from 2x faster overall.

FWIW, LuaJIT runs everything 2-4x faster for me, and more so with long-running loops. Lua's semantics are simpler, though, so it's easier to determine which optimizations are applicable when. (The only downsides for LuaJIT are that it's only for i386, and that it uses about a third more space.)

50% faster on a long running for loop is not the same as 50% faster baseline from CPython... At least from my knowledge PyPy's RPython is not faster than CPython.

50% faster... on selected benchmarks. Run pypy through the unladen-swallow performance benchmarks (very real-world tests) and get some more accurate numbers.

Why do people always benchmark JVM-ported languages to be lot faster than the main version, but in actual production use they turn out to be much slower?