Darn! It turns out that I was naive and jumped to the first conclusion at hand. The $Application object does admit 8 bit chars, it was character 0x00 the one to blame! After running tye's code and looking carefully at the output I found out that $Application returns strings that are truncated at the first appearance of character 0x00. This char seems to be a string terminator for whatever language the object was coded with (a vague memory of having learned this at college comes afloat).

After running a modified version of tye's code (below) I was able to further push the thawing time for my 17K elements hash to 0.15s. This and the time it takes to move memory from $Application gives me a decent 0.26s total, which is 10x better than going to the database and well enough for our expected site traffic.

One big and a respectful bow towards tye!

{
  my( $quote7, $quote8, $zero, $one, %quote, %unquote );
  BEGIN {
    $quote7= pack "C", 0x7e;        # Any 7-bit char.
    $quote8= pack "C", 0x7f;        # Any _other_ 7-bit char.
    $zero = pack "C", 0x00; 
    $one = pack "C", 0x01; 
    @quote{ $quote7, $quote8, $zero }=
      ( $quote7.$quote7, $quote7.$quote8, $quote7.$one );
    %unquote= reverse %quote;
  }
  sub strip8 {
    my( $bin )= @_;
    $bin =~ s#([$quote7$quote8$zero])#$quote{$1}#go;
    return $bin;
  }
  sub restore8 {
    my( $str )= @_;
    $str =~ s#([$quote7$quote8].)#$unquote{$1}#gos;
    return $str;
  }
}
[download]