Thank you very much for your proposals.

There are no additional @INC paths (actually I can see their traversal with strace and it's instant) and I am also testing on both vmware and real machines.

About the I/O stats - the only thing I noticed is that because of that behaviour one of the CPUs on the "problematic new OSes" always get burnt up to 100%.

Additional info: The older OSes are behaving just like the perl file was bytecoded. Actually if I bytecode it, it starts to check syntax for the same amount of seconds on the new OSes. Unfortunately, bytecoding is not a production-stable variant to use ...