Well, I've narrowed it down to ......

...

  langof()
[download]

in

  Lingua::Identify
[download]

Called from

Text::Compare
[download]

If I set the lang to 'en' in Text::Compare's get_words() subroutine, it runs at a steady ~24MB on Linux. Of course, the metric is now different, since it's not "correctly" (as if it ever really was) identifying the languages, but for my app that's not really a problem, since I don't care. As long as it doesn't skew the results badly, - and since this isn't the only metric I'm using, it's not looking that bad.

It's main effect is to cause the matches I get back to indicate more similarity than before. And, actually, I'd rather have more false positives - so it doesn't hurt too much.

I'll talk to the Lingua people next I think now....

Thanks for everyone's suggestions.

Kind regards

Derek.