No matter how long I run it for, it is remarkably consistent here with no more than 1 or 2% variation:

C:\test\perl-5.14.0-RC1>perl
use Benchmark qw[ cmpthese ];;
print $];;
$str = '123456789';
cmpthese -10, {
 a=>q[ my ($a,$b) = $str =~  m/(23)[^8]+(8)/g; ],
 b=>q[ my ($a,$b) = $str =~  m/(23)[^8]+(8)/; ],
 c=>q[ my ($a) = $str =~  m/(23)/g ],
 d=>q[ my ($a) = $str =~  m/(23)/; ],
};;
^Z
5.014000      Rate    b    a    d    c
b 357543/s   -- -15% -33% -45%
a 422192/s  18%   -- -21% -35%
d 535621/s  50%  27%   -- -18%
c 653518/s  83%  55%  22%   --
[download]

One difference of note is that I'm using Window rather than your Linux. Your results reflect ikegami's, who I believe was also using Linux. Perhaps the OP is on Windows?

The 'usual suspect' for performance differences a between those two is memory allocation, but there is none worthy of note here. Indeed, there appear (as you would suspect), to be no calls at all into the OS during benchmark.

Since were both on 64-bit intel hardware, that doesn't seem likely as a cause. Which pretty much leaves only compiler differences, with teh tentative conclusion that with the /g switch enabled, the Windows takes a code path that causes (or allows) the MSC compiler to generate a particularly efficient piece of code somewhere.