First off, I don't like that rand there - it makes things much more difficult to replicate. So, after running your test as-is a few times, I went and just changed "rand(1000)" to "1000" and reran it. I got slightly more consistant answers when the strings where consistantly sized.

At no time have I had pack outperform join. Although my CPU seems to be a bit slower than yours ... I never got over 100 runs per second.

As-is:
       Rate pack join
pack 80.0/s   -- -15%
join 93.7/s  17%   --

Always 1000:
       Rate pack join
pack 46.0/s   --  -8%
join 50.0/s   9%   --

x $i rather than x 1000 or x rand(1000):
       Rate pack join
pack 74.0/s   -- -14%
join 86.0/s  16%   --
[download]

Since it seems that the longer the strings, the closer they were, I tried 10,000.

       Rate pack join
pack 5.66/s   --  -5%
join 5.94/s   5%   --
[download]

It doesn't seem that I can get pack outperforming join on this issue. And that's probably despite being utf-8 aware...

Tests above were run on a threaded perl 5.8.7 on Linux