I would say that anything under 5% is noise and is almost always safely ignorable. (Obviously, 1% for Amazon, EBay, or Google isn't ignorable, but I don't think any of us have that kind of demand.) Of course, this is a rule of thumb with the appropriate caveats about such.

Furthermore, I don't think benchmarking variable name length isn't worthwhile, for the simple reason that Perl itself is a poor choice for execution speed. If you have optimized your Perl code to the point that the only thing you can think to do is reduce the size of your variable names and you still need to optimize further, then you need to look at rewriting some of your Perl in XS and/or C. If that doesn't help, rewrite the rest of your Perl in C and optimize some of your C to ASM. If that doesn't work, you need bigger servers.

In other words, micro-optimizations like that are cases of premature optimization and should be avoided. But, then again, you already knew that. :-)