xen vs os

while doing my previous tests with _datafile_write_errors_crash_instance with  on a paravirtualized  linux (centos 5.7 64bit)  under xen server 5.5 I had some unexpected behaviour, basically

I removed the datafile and oracle, or better the Dbwriter didn’t see it at all. I even could continue to create tables and data in that datafile.

i went to /proc/PID_OF_THE_DBWR/fd
and there did …

ls -al

lrwx—— 1 oracle oinstall 64 Mar  2 14:44 260 -> /u02/MYDB/undotbs01.dbfsnipped …lrwx—— 1 oracle oinstall 64 Mar  2 14:44 261 -> /u02/MYDB/users01.dbf (deleted)

and it was really deleted, but somehow it wasn’t picked up. I was a bit astonished by this behavior and decided to redo the exercise on a non virtual linux Centos 5.6
there the same action almost immediately got picked up by oracle  and the datafile was brought offline or the instance crashed depending on how you set _datafile_write_errors_crash_instance.
Pretty bizar. 
an strace of the delete  on both systems revealed the same os calls.

My colleagues and I discussed this and our best guess is the Xen Server “optimizes” the data access an causes this behaviour.

We will test this next year with a pure virtualized linux ….
keep you posted

Leave a Reply