The startup overhead is caused by the idea to have the stdlib in perl6 source code (similar to mature languages like common lisp), and that the compiler will be good enough to produce good enough code. But the sigs of all methods need to be dumped somehow into the binary, and this done beyond naive. Emacs or Common Lisps solved this problem pretty well, via dumping native image of the compiled code, java has a good native class layout, but rakudo is just slow.
The nqp overhead is when you look at the compiled moar code and compare the bytecode to a fast vm. Its still too bad code. moar itself is fine, but the rakudo and nqp layers not. Same for the new jit template idea, which does not scale. At all. There are so many people with so many bad ideas who are not willing to listen to more experienced dev's, so I stopped caring also. It's not worth it, esp. after the parrot debacle. But perl6 people are generally very nice, competent and open, in total opposite to p5p.
How so?