Re: Fast object checkpoint

On 09/22/2006 08:02:52 PM, Mladen Gogala wrote:
> What is a "fast object checkpoint"? I have script which has to update ~60M rows.
> I had to kill it and re-start it. The new incarnation did not wait for a bunch of
> TX locks, it was waiting for the following event:
>
> ENQ: KO - fast object checkpoint
>
> Surprisingly enough, this lock was held by the CKPT process, not PMON, as I've
> expected. Very few undo blocks were actually consumed in v$undostat. My question
> is what is "fast object checkpoint" and is there a new method of reclaiming resources
> held by murdered transactions? Is it documented anywhere?
>
> --
> Mladen Gogala
> http://www.mgogala.com
>

I found the description of this new feature in the following document:

www.oracle.com/technology/deploy/performance/pdf/twp_perf_database%20performance%20with%20oracle10gr2.pdf

"Prior to Oracle Database 10g administrators could specify the expected crash recovery time (MTTR) by setting the value of a checkpoint-related initialization parameter (FAST_START_MTTR_TARGET). They could do so by using the MTTR advisory, which helps predict the number of physical writes that would arise with different MTTR target values. With Oracle Database 10g, the database can self-tune checkpoints activity to achieve good recovery times with low impact on normal throughput. With automatic checkpoint tuning, Oracle Database takes advantage of periods of low I/O usage to write out data modified in memory to the data files without adverse impact on the throughput. Consequently, a reasonable crash recovery time can be achieved even if the administrator does not set any checkpoint-related parameter or if this parameter is set to a very large value. Another enhancement done in the second release of Oracle Database 10g dramatically improves the performance of object-checkpoint requests issued for objects accessed through direct path reads, a situation that can occur with parallel query. Before an object can be accessed through direct path reads, dirty buffers of the object must be written to data files on disk via an object-checkpoint request. Prior to Oracle Database 10g Release 2, the checkpoint request is handled by issuing a checkpoint for the tablespace the object belongs to, writing out all the dirty buffers for the entire tablespace. Since a large number of objects may reside in the same tablespace, this implementation may cause large number of unnecessary disk writes. With the new release, a checkpoint request for a target object will only write out the dirty buffers of that object, without incurring any additional writes for the dirty buffers of other objects"

Of course, this is a very important new performance feature. Many people have noticed and inquired about the "KO locks", queried v$lock_type and didn't investigate any further, but this dramatically changes the way the database functions. It also dramatically impacts performance consideration as a big buffer cache in which large parts of a big table can be cached can cause a serious I/O contention and a lock contention. I'm looking for a mechanism to turn off this new behavior, at least until the next patch version, if not until the next major version.

-- 
Mladen Gogala
http://www.mgogala.com

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