The big takeaway was this: Most of the things you think you know about Java performance are really hard to quantify. Write the best code you can. Profile it to find the bottlenecks. Remove the bottlenecks. Rinse and repeat.
A couple of other takeaway points here:
- Don't forget to upgrade every so often. JDK6, for example, runs code compiled for JDK4 just fine, and you get a big performance boost. For free.
- Memory allocation is pretty cheap. Know that there are three garbage collectors, and know which one is best, when. Read the Java memory management white paper.
- Hotspot can do lots of things to optimize your locks, but it often doesn't. Having said that, the bottlenecks in concurrent code usually have to do with contention, not synchronization overhead. Use low contention data structures like ConcurrentHashMap to avoid contention, but don't try to get clever with locking.
- Contrary to popular belief, thread-per-request synchronous servers are often faster than selector-based asynchronous servers.
The Java Performance Myths slides are here.