I observed that the nanos are calculated in slices, mean if measured you will see i.e.100/100/100/200 so it looks like that it is not real-time, but this is a bit in the past with earlier Java 1.6.
The performance I've seen was around 1.000 calls/sec with a single thread client on a blade system with Linux (I can't remember the exact configuration). If I find time I'll run a simple test :)