Let me start with a lesson I learned recently: with your code, you are not using the $a and $b you declared at the top. The code executes without warning because $a and $b are "reserved" variables in Perl (used in sorting). If you change the variables to something else (capitalize them), you get:

sub traditional {
    my $C = $A;
    $A = $B;
    $B = $C;
}

sub non_trad {
    ($B, $A) = ($A, $B);
}

Global symbol "$A" requires explicit package name at ./bench.perl line
+ 12.
Global symbol "$B" requires explicit package name at ./bench.perl line
+ 13.
[download]

You should "use vars" to allow Benchmark to use the variables you want it to use.

This doesn't change the results, though, since you were (unknowingly) using "real" variables, anyway.

Here are some more possibilities, just to play around:

#!/usr/bin/perl -w

use vars qw/$A $B @Arr/;

use strict;
use Benchmark;

$A = rand(10);
$B = rand(10);
@Arr = ($A, $B);

sub traditional {
    my $C = $A;
    $A = $B;
    $B = $C;
}

sub non_trad {
    ($B, $A) = ($A, $B);
}

sub array_slice1 {
    @Arr = @Arr[1,0];
}

sub array_slice2 {
    @Arr[1,0] = @Arr;
}

timethese (100000, {
    "traditional" => \&traditional,
    "not traditional" => \&non_trad,
    "array1" => \&array_slice1,
    "array2" => \&array_slice2,
});

Results:
Benchmark: timing 100000 iterations of array1, array2, not traditional
+, traditional...
         array1:  4 wallclock secs ( 3.98 usr +  0.00 sys =  3.98 CPU)
         array2:  3 wallclock secs ( 3.27 usr +  0.00 sys =  3.27 CPU)
not traditional:  2 wallclock secs ( 2.72 usr +  0.00 sys =  2.72 CPU)
    traditional:  2 wallclock secs ( 1.81 usr +  0.00 sys =  1.81 CPU)
[download]

Russ
Brainbench 'Most Valuable Professional' for Perl