I personally believe that Oracle instrumented its code for kernel optimization, not for DBAs or application developers. An obvious clue is the names given to describe the timing, such as db file scattered read to relate non-buffer cache multiblock reads and latch free to signify latch sleep time. No DBA would have come up with names like these!

Oracle has never guaranteed it has completely instrumented it code. This leaves open the possibility for unaccounted for time. If you are counting each millisecond and come up short, one of the possible reasons for this is that Oracle simply did not instrument a piece of code you ran. For this, and many other reasons, I don't expect to account for every little bit of time in my analyses. This may seem like a problem, but I don't believe it is.

Oracle performance analysis is about solving real-life production performance issues. In all my years of using Oracle's instrumentation, Oracle has not misled me in my analysis. The OraPub 3-circle analysis has confirmed and strengthened what Oracle's instrumentation has indicated. This is why I don't get uptight about unaccounted time. My job is to fix real performance issues, not to find the missing 50 ms!