I spend some time trying to reproduce the mdr hotspot, using the Cinco deployment with 50 beans and 2000 methods, and what I found was that AOP was the primary bottleneck, hardly anytime was spent in mdr.
Ugh, AOP is using CopyOnArraySets to hold Methods. That means Method.equals on the entire data structure for every method before copying the whole thing.
One of the AOP bottlenecks:
https://jira.jboss.org/jira/browse/JBAOP-716