I just presented to the Toronto Perlmongers (badly) about how I did this. The short answer is that I used a daemon that spawned a bunch of kids to do some work. Requests arrived at the mod_perl request handler asynchronously, and were handled the same way.

I can highly recommend IPC::Run, with the following caveats:

• Specify a timer when setting up the harness. Failing to do so seems to cause the start method to not return.
• Don't forget to use either pump or pump_nb after sending a command or before looking for the response to a command -- it acts like a yield (remember non-preemptive multi-tasking?).
• Do your IPC::Run testing with a simple, predictable program, to make sure that you're figured out the IPC::Run functionality before you try to run the actual Production application.

I hope to post soon on How Not To Make A Presentation to a Local User's Group, as well as Some Pointers on Using IPC::Run. The Toronto Perlmongers will know what I'm talking about.