I believe the original benchamrk code as posted is broken. Thanks to Aighearach for questioning my earlier results and leading me along this path.

Consider this code:

my $now = 8;
my %url = ( monday => { @{[map(($_,1), (1..1000))]} } );
timethese(0, {
  Grep  =>      'Grep();',
  Grep2 =>      q{(sort grep {$_ <= $now} keys %{$url{"monday"}})[-1];
+},
  Grep3 =>      sub{(sort grep {$_ <= $now} keys %{$url{"monday"}})[-1
+];}
});
sub Grep{
  $now = (sort grep {$_ <= $now} keys %{$url{'monday'}})[-1];
}
[download]

The output from that:

Benchmark: running Grep, Grep2, Grep3, each for at least 3 CPU seconds...
      Grep:  4 wallclock secs ( 3.00 usr +  0.02 sys =  3.02 CPU) @ 219.21/s (n=662)
     Grep2:  4 wallclock secs ( 3.14 usr +  0.01 sys =  3.15 CPU) @ 356475.56/s (n=1122898)
     Grep3:  4 wallclock secs ( 3.30 usr +  0.04 sys =  3.34 CPU) @ 213.47/s (n=713)

shows that something is definitely up, the "inline sub" and "called sub" take the same time, but the "inline quoted" version is unbelievably faster. I think it has something to do with the inline quoted version not really working the way the original poster intended (I think its some sort of quoting/scoping problem, but I'm not sure exactly what).

I have rerun my benchmarks with all the routines as "inline subs" and the results follow:

dataset size	Grep	Max	Ternary
100	2440.20/s	2511.96/s	2495.22/s
1000	207.32/s	222.81/s	216.31/s
10000	15.31/s	16.77/s	16.61/s

And here is the code I used to run the test:

   my %url;
   my $now = 8;

   my @size=(100,1000,10000);
   foreach(@size){
     my $size=$_;
     print "size=$size\n";
     %url = ( monday => { @{[map(($_,1), (1..$size))]} } );
     timethese(0, {
        Grep    =>      sub {(sort grep {$_ <= $now} keys %{$url{"mond
+ay"}})[-1];},
        Ternary =>      sub {($now < $_ && $_ < 8 ? $_ : $now) for key
+s %{$url{"monday"}};},
        Max     =>      sub {foreach ( keys %{$url{"monday"}} ) { $now
+ = $_ if $_ > $now };}
        });
     undef %url;
   }
[download]