Re: RAC partitioning idea bounce
From: goran bogdanovic <goran00_at_gmail.com>
Date: Sat, 23 Nov 2013 14:00:59 +0100
Message-ID: <CAGyPXK6a4PsYSri2SwF+K_5k76=8f9FCfjgbtjP5qKJK7gYQVg_at_mail.gmail.com>
right, of course ... in case of single instance database, transactions are generated in a local instance only ... whereas in RAC database, transactions can be initiated in any instance hence possibly creating all sorts of contentions.
Contention can be on undo header blocks, undo blocks, index blocks, table blocks, block headers, sequences etc.
As per description in OP, contention on indexes will be unavoidable ... i.e as many processes will be concurrently accessing/modifying index blocks, so hash-partitioning index is only one measure. On top of it, further investigation in what other sorts of segments are affected by contention will reveal what other measures are necessary e.g. redesign table to use hash-partitioning, reduce the number of indexes on the table, index compression etc. ...
But localized access sounds even better approach to me in such case ... if load performance is not satisfying, traditional strategies for speeding-up local load routine may probably give better overall result than spreading the load job on all instances in cluster.
Date: Sat, 23 Nov 2013 14:00:59 +0100
Message-ID: <CAGyPXK6a4PsYSri2SwF+K_5k76=8f9FCfjgbtjP5qKJK7gYQVg_at_mail.gmail.com>
right, of course ... in case of single instance database, transactions are generated in a local instance only ... whereas in RAC database, transactions can be initiated in any instance hence possibly creating all sorts of contentions.
Contention can be on undo header blocks, undo blocks, index blocks, table blocks, block headers, sequences etc.
As per description in OP, contention on indexes will be unavoidable ... i.e as many processes will be concurrently accessing/modifying index blocks, so hash-partitioning index is only one measure. On top of it, further investigation in what other sorts of segments are affected by contention will reveal what other measures are necessary e.g. redesign table to use hash-partitioning, reduce the number of indexes on the table, index compression etc. ...
But localized access sounds even better approach to me in such case ... if load performance is not satisfying, traditional strategies for speeding-up local load routine may probably give better overall result than spreading the load job on all instances in cluster.
On Fri, Nov 22, 2013 at 7:31 PM, mvshelton <mvshelton_at_chartermi.net> wrote:
> Did they specify what the contention is? If not I would first narrow down
> what is the contention wait event. It could be tablespace, sequence, and
> what everyone is talking about segment contention. Without knowing the
> exact wait event partitioning may or may not solve your issue.
>
>
> Sent via the Samsung Galaxy S™ III, an AT&T 4G LTE smartphone
>
>
>
-- http://www.freelists.org/webpage/oracle-lReceived on Sat Nov 23 2013 - 14:00:59 CET