Should be most recent timing for each major subsystem for each type of call. E.g.,

PUT: 10ms (locking), 10ms (datacontainer), 20ms (RPC), 30ms (CacheStore).
GET: 0ms (locking), 5ms (datacontainer), 0ms (RPC), 10ms (CacheStore).

etc.

Could be implemented as a simple ringbuffer, in a specific component (TimeBudgetMonitor?) just storing the most recent N calls and that's it. Cheap to capture, cheap to store.

This data could then be made available via a JMX operation on TimeBudgetMonitor. This is extremely valuable for tuning and debugging perf issues.