A streaming parser doesn't seem like a very threadable kind of thing insofar as you're talking about walking a document tree in parallel. I think you'd end up doing a great deal of work trying to rig something where, for example, you'd have dedicated threads for processing particular node types.

I think I'd be more inclined to try to break the raw input into separate chunks that could run in parallel processes, something your 28 processor machine could do very well. So, maybe rather than reengineering your existing code to use threads, you could write a frontend script to divide whole units of work and then fork off processes.

Hard to say much more without concrete examples.