Re: System statistics
From: Christian Antognini <Christian.Antognini_at_trivadis.com>
Date: Mon, 19 Aug 2013 06:15:15 +0000
Message-ID: <43D812FCDBAB8948B2F14F61C4B1CD0A74254368_at_smxc002.trivadis.com>
Hi Tony
Date: Mon, 19 Aug 2013 06:15:15 +0000
Message-ID: <43D812FCDBAB8948B2F14F61C4B1CD0A74254368_at_smxc002.trivadis.com>
Hi Tony
Even though you probably know what I think about it, here is a short feedback...
> Given our tolerance for cardinality errors why are we so concerned about
> accurately quantifying the size and performance of multi-block reads?
- I see execution plans flipping for much less than a factor of 2 or 3.
- You increase the likelihood to see features like cardinality feedback kicking in.
> Why not just delete system statistics and set DB_FILE_MULTIBLOCK_READ_COUNT
> to 0 (implying the maximum possible I/O size) and be done with it?
- Even though you delete them they will be automatically re-gathered. As of 10g it is simply not possible to run a database without system statistics.
- The auto DB_FILE_MULTIBLOCK_READ_COUNT is not good for at least two reasons:
- Suboptimal performance is possible
- When MEMORY_TARGET or SGA_TARGET are in use the value can change at every instance bounce. As a result, execution plans can change when you bounce an instance.
Just my two cents...
HTH
Chris Antognini
Troubleshooting Oracle Performance, Apress 2008 http://top.antognini.ch
-- http://www.freelists.org/webpage/oracle-lReceived on Mon Aug 19 2013 - 08:15:15 CEST