Re: AIX Mirror Write Consistency

Comments inlined -

Jonathan Lewis wrote:

> Is this the same as 'dirty region logging' on other
> systems - a very dense (1 bit = 32K type of thing)
> map which tells the O/S the status of mirrored writes ?
>

It sounds like exactly that...

> What if -
> A database block is written to one side of the mirror
>

But I don't think Oracle will accept that as written unless both sides have beenwritten to..

> The system crashes
>
> The system is restarted, and recovery run
>
> The recovery process reads the mirror which has
> been written to correctly and decides it does not
> need recovery.
>

I would think that Oracle would roll back this transaction, and thereforeit would re-write correct information..

> At a later date, when the database has been running
> for some time, the block is read from the other side
> of the mirror ?
>

Is it really that random? There is the concept of a primary isn't there? Also,I would think that the rollback would have re-written this information, so that the
mirrors would be in sync again anyhow. I think it all depends on wether or not Oracle
waits for writes to go thru to the mirror or not before considering it a committed transaction. If it doesn't, that I can see problems. But the sync deamon must be doing something..

> How does the O/S, or Oracle, determine that you have
> one block on one side of the mirror that is correct and
> the other isn't ?
>

Good question.. but I think this is only a danger if Oracle accepts it's transactionas comitted if it's only one one side of the mirror. If the loop is busy, and Oracle
decides to read from the mirrored disk instead, and it has a half comitted transaction
written on it, I could see an issue of this sort coming up, but I'm not sure that's possible.

>

Also, only for IBM RS/6000 AIX, there is note on metalink # 1030317.6 that insiststhat if you are using parallel server, you MUST have it turned off.

" Logical volumes used as raw character devices for concurrent access from two
 nodes must have the Mirror Write Consistency and Bad Block Relocation  attributes set to no. "Why would that be ok for parallel server, and not single server. I would think the
potential for problems of the sort you've mentioned would be even greater on parellel server.
Maybe this is a platform specific issue?

• Dc.

> --
>
> Jonathan Lewis
> Yet another Oracle-related web site: http://www.jlcomp.demon.co.uk
>
> Doug Cowles wrote in message <38627CC7.D06EA4F8_at_bigfoot.com>...
> >Does anyone have any experience with turning off MWC (Mirror Write
> >Consistency)
> >on AIX with Oracle? My guess is it is ok. A note on metalink said it
> >is a requirement
> >for parallel server, and a small document I found (which I could post)
> >- says the following:
> >
> >"Data Integrity :
> >You can only have problems (by turning it off) if you are using a
> >program which can read
> >data for which no completion stats was given for the previous write.
> >
> >Even without mirroring, if a system crash occurs whilst a write is
> >outstanding there is not a lot you can say about the data area on the
> >disk when the system comes back. It could be old data or new data or
> >some mixture of some other data entirely.
> >
> >Relational databases use a log file to allow them to recover from a
> >crash so, proviced the log file algorithm allows for this condition,
> >there should be no problem. You should check
> >with your app vendor to see if there recovery algorithm for a system
> >crash can cope with mirror write consistency turned off. I suspect that
> >most can. "
> >
> >Oracle does not even finish a commit until it has received that write
> >has completed no?
> >So it should be ok I think. Unfortunately, I will need an official
> >answer from Oracle I
> >think. Turning it off can give a huge performance boost I read.
> >
> >Any thoughts?
> >- Dc.
> >
Received on Thu Dec 23 1999 - 17:24:26 CST