----- Original Message ----- From: "Gerry Miller" <gerry@xxxxxxxxxxxxxxxxxxx> To: <jonathan@xxxxxxxxxxxxxxxxxx> Cc: <oracle-l@xxxxxxxxxxxxx> Sent: Tuesday, June 19, 2012 7:07 AM Subject: Re: High CPU usage | It is quite a lightweight query of sys.all_tab_cols and obj$ but the | reason I chose it was that it had exactly the same buffer gets in both | files and the same type and number of waits events. I was trying to | avoid any red herrings. The key question, though is whether the number of buffer gets seems reasonable for the query. If we can't trust the number of buffer gets reported then the difference in the CPU usage becomes a secondary problem. | | Am I correct in assuming that any such latch or mutex problem activity | would manifest itself as a session wait event? | Not necessarily - latches spin before sleeping and mutexes spin - both of which take up CPU. Normally we would expect a relatively short spin followed by a sleep, of course, but your tkprof output is anomalous so we need to think about possible reasons for unexpected behaviour (i.e. bugs). On an O/S-related variant of the same theme - is your regression testing on a new (or upgraded) platform to the one you've otherwise been using. Is it possible that some sort of NUMA side effect is making memory access very slow and causing extreme CPU stalling when accessing latches or mutexes ? Can you compare the O/S CPU reporting with the Oracle CPU reporting for one of these anomalous tests. Regards Jonathan Lewis -- //www.freelists.org/webpage/oracle-l