Here's the thing. It's not slower. It's actually faster, but your benchmark wasn't constructed properly. You used a constant, which I think Perl was optimizing away. Also, you were burying sprintf within a relatively expensive function call. Take a look at this:

use Benchmark qw[cmpthese];

sub one {
        sprintf("%09d",@_);
}

sub two {
        substr("000000000$_[0]",-9);
}

cmpthese(100, {
        one => sub { for(0..10000) { $_ = one($_) } },
        two => sub { for(0..10000) { $_ = two($_) } },
        tre => sub { for(0..10000) { $_ = sprintf("%09d", $_) } },
});

__DATA__
Benchmark: timing 100 iterations of one, tre, two...
       one: 12 wallclock secs (10.63 usr +  0.00 sys = 10.63 CPU) @  9
+.41/s (n=100)
       tre:  3 wallclock secs ( 3.38 usr +  0.00 sys =  3.38 CPU) @ 29
+.59/s (n=100)
       two: 12 wallclock secs (11.08 usr +  0.00 sys = 11.08 CPU) @  9
+.03/s (n=100)
      Rate  two  one  tre
two 9.03/s   --  -4% -69%
one 9.41/s   4%   -- -68%
tre 29.6/s 228% 214%   --
[download]

You're correct in your assumption that substr is faster, so long as the only number you need to process is "213" and you bury it in a subroutine call. You'll have to admit that in the real world users tend to ask for a little more functionality than that.

It would seem that sprintf wins by a knockout.