Re: Update Statements and Exadata

From: Martin Berger <martin.a.berger_at_gmail.com>
Date: Sun, 9 Dec 2012 14:45:37 +0100
Message-ID: <CALH8A92bag80vya2y1tsSOt1rjvVg7UCpTyh=+rHNprFmPLU_A_at_mail.gmail.com>



I'd say, for an update you allways need blocks in buffer cache, whereas for a select Cells can deliver prepared data already - so there is no block structure anymore.
Now my guesswork starts:
Someone at Oracle decided "if only one process is running the update, it need to get all the blocks (for update) into buffercache at all - so don't let double the work, which would be: first transfer data-only (serial_direct) - afterwards transfer the blocks (sequential read?) For parallel processing the thing might be different: the parallel processes wich consume the data might be delivered via offloading, but those parallel processes which do the update itself has to go via buffer_cache/conventional.

Again, just a guesswork; I'd be hope to get detailled insights myself :-)

Martin

On Sun, Dec 9, 2012 at 2:32 PM, coskan gundogar <coskan_at_gmail.com> wrote:

> That is alright but what is changing with the undo management when issued
> parallel so it can be offloaded as if i just select instead of update
> regardless of degree it is always ofloading
>

--
http://www.freelists.org/webpage/oracle-l
Received on Sun Dec 09 2012 - 14:45:37 CET

Original text of this message