2 things can help you here: you can use start_tag_handlers, which are called after the start tag of the element has been parsed (and the element object has been created, empty at that point). This also means taht the UML:Attribute element will get completely parsed before the UML:Class, but when you're in their handler the opening tag of UML:Class has already been parsed. If you don't use twig_roots but regular twig_handlers the element exists, it's an ancestor of the UML:Attribute element, and it's attributes are already available. If space is a problem, you cant sprinkle purge call to taste.

So in your case I would write something like (untested):

use strict;
use warnings;

use Data::Dumper;
use XML::Twig;

my $twig = XML::Twig->new(
  start_tag_handlers => { 'UML:Class' => \&ulm_class, }, 
  twig_handlers      => { 'UML:Class' =>  sub { $_[0]->purge }, # purg
+e at the end of each section,
                          'UML:Attribute' => \&uml_attr,
 } );
$twig->parsefile( 'testfile.xmi' );


sub uml_class
{
    my ( $twig, $section ) = @_;

    print "data for class:\n";
    print "   name = ", $section->att( 'name' ), "\n";
    print "   xmi.id = ", $section->att( 'xmi.id' ), "\n";
}


sub uml_attr
{
    my ( $twig, $attr ) = @_;
    # if you need the class id, it's in $attr->parent( ''UML:Class')->
+attr( 'xmi.id')
    $attr->print;

    print Dumper( $struct );
    # parse the block and extract the data elements

    $twig->purge;
}
[download]

Does it make sense? It probably doesn't matter that much if your files are small, but it feels better to parse only once each section.