RE: Performance Problem.

> Now lets go to the cool sides of real Unixes. :) In Solaris
> 9, you can see what is TLB misses overhead using "trapstat
> -T". It will show % of time lost due to TLB miss. Works on
> SPARC but I'm not sure it will work on x86 as well.

Trapstat is not implemented on x86. But you can use prstat -m and look at the TRP, TFL and DFL columns for trap, text page fault (executable) and data page fault times for a process (or thread with -L option).

Regarding original posters observation that semaphore operations did increase during the slowdown - this *may* be a symptom of a jump in latch contention (as contending processes need to go to sleep and then be woken up etc). This could be caused purely due the workload (inefficient query being executed or CPU heavily overloaded) or it could be due something happening at low-level in Oracle. Shared pool fragmentation and large allocations come to mind first, but this could be anything else, and we can't be sure given that we don't have any data to work on.

--
Regards,
Tanel Poder
http://blog.tanelpoder.com


--
http://www.freelists.org/webpage/oracle-l