It seems you have misunderstood my posting, perhaps it was a little bit short on information.

I did not run NYTProf once or twice to see that factor 30. I ran NYTProf after every little change, to see if that change was accelerating my script or not, and to find the next hot spot in the code. Some changes had little or no effect, some even slowed down. Most changes gained only some percents. The two big winners were the traversing code and XML::Twig.

Removing the traversing code already accelerated my script. Some smaller steps followed, but then the next problem became quite obvious: XML::Twig burned a lot of time when creating new elements and attributes. I replaced XML::Twig with XML::LibXML instead of hacking XML::Twig simply because it was easier than creating a private branch of XML::Twig fine-tuned for my special case. It was a reversable test, because I had the latest version of the script using XML::Twig in SVN. If XML::LibXML had been as "slow" as XML::Twig, I could have simply issued a "svn revert" and could have continued by optimizing a private branch of XML::Twig. This replacement was mostly a quite simple search-and-replace operation for the different class and method names, but it gave an instant performance boost of several hundred percent. So I commited that version and never thought about going back to XML::Twig. The one thing I lost was the really useful "csv" indent style available in XML::Twig output, but I added that back later using some perl code. After replacing XML::Twig with XML::LibXML, I found and reworked some other routines wasting some time, but compared to the tree traversing code and XML::Twig, they were nearly insignificant. Now, the script is sufficently fast and spends most time in code that I can not optimize further (Perl opcodes and libxml xsubs).

Alexander