Try a Schwarzian transform based code:

use strict;
use Data::Dumper;

 my %data;

  #-- original sample
  $data{1} = {'a' => 10, 'b' => 5, 'c' => 20 };
  $data{2} = {'a' => 90, 'b' => 15, 'c' => 40 };
  #--

 my @Top =
        sort { $a->[2] <=> $b->[2] }
        map {
                my $h = $_;
                map {
                        [ $h, $_, $data{$h}->{$_} ]
                  } keys %{$data{$h}};
         } keys %data;

  print Dumper(\@Top);
[download]

On Schuartzian transformation you receive the original data format sorted, but you don't need it here.

I don't see my last sugest here...

One more Schwarzian transform based code..