I cobbled this together based on a post on clpm from several years ago.

tr beats out s by between 4 and over 10 to 1 if I have interpreted this correctly.

use strict;
use Benchmark;
 
 my $string1 = join '', map rand() < 0.4 ? 'x' : ' ', 1 .. 100;
 my $string2 = join '', map rand() < 0.4 ? 'x' : ' ', 1 .. 10000;


 printf "Case 1 deletes %3d of %4d characters\n",
   $string1 =~ tr/ / /,
   length $string1;
 printf "Case 2 deletes %3d of %4d characters\n",
   $string2 =~ tr/ / /,
   length $string2;
 
 timethese( 10000, {
   tr1 => sub { my $a = $string1; $a =~ tr/\n\t //d},
   s1  => sub { my $a = $string1; $a =~ s/\s//g},
   tr2 => sub { my $a = $string2; $a =~ tr/\n\t //d},
   s2  => sub { my $a = $string2; $a =~ s/\s//g},
 });
 __END__
Case 1 deletes   7 of   10 characters
Case 2 deletes 592 of 1000 characters
Benchmark: timing 100000 iterations of s1, s2, tr1, tr2...
        s1:  2 wallclock secs ( 1.13 usr +  0.00 sys =  1.13 CPU) @ 88
+339.22/s (n=100000)
        s2: 54 wallclock secs (52.58 usr +  0.00 sys = 52.58 CPU) @ 19
+01.68/s (n=100000)
       tr1:  0 wallclock secs ( 0.30 usr +  0.00 sys =  0.30 CPU) @ 33
+2225.91/s (n=100000)
            (warning: too few iterations for a reliable count)
       tr2:  3 wallclock secs ( 4.34 usr +  0.00 sys =  4.34 CPU) @ 23
+062.73/s (n=100000)

Case 1 deletes   5 of   10 characters
Case 2 deletes 615 of 1000 characters
Benchmark: timing 100000 iterations of s1, s2, tr1, tr2...
        s1:  1 wallclock secs ( 0.97 usr +  0.00 sys =  0.97 CPU) @ 10
+2880.66/s (n=100000)
        s2: 54 wallclock secs (54.22 usr +  0.00 sys = 54.22 CPU) @ 18
+44.37/s (n=100000)
       tr1:  1 wallclock secs ( 0.31 usr +  0.00 sys =  0.31 CPU) @ 32
+1543.41/s (n=100000)
            (warning: too few iterations for a reliable count)
       tr2:  5 wallclock secs ( 4.39 usr +  0.00 sys =  4.39 CPU) @ 22
+753.13/s (n=100000)

Case 1 deletes  54 of  100 characters
Case 2 deletes 5999 of 10000 characters
Benchmark: timing 10000 iterations of s1, s2, tr1, tr2...
        s1:  0 wallclock secs ( 0.54 usr +  0.00 sys =  0.54 CPU) @ 18
+484.29/s (n=10000)
        s2: 54 wallclock secs (52.71 usr +  0.01 sys = 52.72 CPU) @ 18
+9.66/s (n=10000)
       tr1:  0 wallclock secs ( 0.06 usr +  0.00 sys =  0.06 CPU) @ 16
+6666.67/s (n=10000)
            (warning: too few iterations for a reliable count)
       tr2:  4 wallclock secs ( 4.30 usr +  0.00 sys =  4.30 CPU) @ 23
+27.75/s (n=10000)
[download]