Re: Using PQ in FTS
- From: "Frits Hoogland" <frits.hoogland@xxxxxxxxx>
- To: "Christian Antognini" <Christian.Antognini@xxxxxxxxxxxx>
- Date: Thu, 29 Mar 2007 03:04:54 -0500
see inline comments
On 3/29/07, Christian Antognini <Christian.Antognini@xxxxxxxxxxxx> wrote:
Frits
> If mixed access (meaning both buffered and direct access)
> is done on objects where DML is happening, it can result
> in buffer busy waits because extents needs to be
> checkpointed before direct access can occure.
Be careful... checkpoints are not performed at extent level...
Up to 10gR1 checkpoints are executed at tablespace level.
Ah! That explains; I've witnessed waits at partitions.
As of 10gR2 checkpoints are executed at segment level (e.g. for a
partitioned table a single partition can be checkpointed).
> Seen a great deal of buffer busy waits in 9.2.0.6/linux where
> concurrent DML queries got downgraded to serial due to the
> parallel automatic tuning.
Do you mean because of parallel adaptive multi user?
Yes, I meant 'parallel_adaptive_multi_user' enabled because of
'parallel_automatic_tuning'.
Point I am trying to make is PX is both extremely powerfull and also
extremely dangerous; even with just doing reads and sorts it requires to
have complete control over the database and the machine(s) the database
lives on. Unexpected extra jobs could lead to performance degradation
instead of performance benefit.
Doing DML takes this even further and can give very bad results if both
buffered and direct DML is done at the same time.
Regards,
Chris
thanks,
frits
Other related posts: