Here is a benchmark:

#!/usr/bin/perl -w

use strict;
use Benchmark;

timethese(1000, {
    'global' => '&global',
    'local' => '&local' });


sub global {
    my %hash;
    for(1..100) {
        for(1..10) { $hash{rand(100)} = 123; };
    }
        %hash=();
}

sub local {
    for(1..100) {
        my %hash;
        for(1..10) { $hash{rand(100)} = 123; };
    }
}
[download]

And it gives:

Benchmark: timing 1000 iterations of global, local...
    global: 14 wallclock secs (13.47 usr +  0.01 sys = 13.48 CPU)
     local: 14 wallclock secs (13.33 usr +  0.02 sys = 13.35 CPU)
[download]

so it seems like there is not much difference.