From what I've understood in the link I've posted in this thread, you have to add audio and MIDI latency, to get the result of the delay you get with all the sound hardware.
One thing I didn't figured out in this topic is that if MIDI latency is almost about one second when audio is on, then the audio latency must be be nearly this high latency as well, unless the amount of MIDI information is huge.
Sorry if I'm picky on importance to know exactly how far the informations are delayed, I've been passing a kind of war in the computer just to get a few milliseconds in the PD processes, by reviewing how the objects [trigger] and [until] are used all along the application.
Another point I've been thinking of that only concern latency and not MIDI nor XP, is that the player must have one meter latency (about 1000/320) if he do play with headphones, so 3ms would be perfect.
With OSS driver on linux the delay goes above 5ms. I've got this result with dividing by two the output of [timer] with the two inlet connected to MIDI in and out.