[Ocfs-users] A couple more minor questions about OCFS and RHE L3

[Sunil Mushran]

Symptoms is a failed write in the logfile which leads to
an immediate db crash. Any io errors on the logfiles
are considered fatal.

ORA-27091: unable to queue I/O
ORA-27072: File I/O error

Derek Suzuki wrote:

>	Sort of a followup...
>	We've been running OCFS in sync mode for a little over a month now,
>and it has worked reasonably well.  Performance is still a bit spotty, but
>we're told that the next kernel update for RHEL3 should improve the
>situation.  We might eventually move to Polyserve's cluster filesystem for
>its multipathing capability and potentially better performance, but at least
>we have a stable, functioning platform for the time being.
>	My DBA still wants to try async mode with OCFS.  We followed your
>recommendations for using dd to recreate the logfiles to be contiguous on
>disk (if we don't do that, they always seem to come up with two
>non-contiguous extents) and are doing some testing.  We were, however,
>wondering what kind of symptoms would appear if we were to trigger the
>non-contiguous aio problem that you have described.  We weren't sure if it
>would result in an immediate failure, or if it would trigger silent
>corruption that we wouldn't notice until much later.
>	We are also considering the possibility of using OCFS for our
>datafiles and raw devices for the redologs.  In theory that ought to
>eliminate the problem entirely.  It's easy enough to reconstruct individual
>logfiles with dd, but I'd be worried that someone might forget to do that
>while building a new database or restoring a dataset from backups.
>
>	Anyway, I wanted to thank you and Sunil again for all of the helpful
>info you've provided us during our RAC deployment.  Circumstances forced us
>to go live sooner than we would have liked, and we needed all the help we
>could get to get everything working.
>
>Derek
>
>  
>
>>-----Original Message-----
>>From: Wim Coekaerts [mailto:wim.coekaerts at oracle.com]
>>Sent: Saturday, March 06, 2004 10:36 PM
>>To: Derek Suzuki
>>Cc: 'ocfs-users at oss.oracle.com'
>>Subject: Re: [Ocfs-users] A couple more minor questions about OCFS and
>>RHEL3
>>
>>    
>>
>>> Next, I saw a Metalink thread which suggests that async I/O is not
>>>supported on OCFS with RHAS 2.1.  It doesn't say anything 
>>>      
>>>
>>about RHEL3.
>>    
>>
>>>We've been using async in our testing with no problems so 
>>>      
>>>
>>far, and plan to
>>    
>>
>>>use it in production unless Oracle feels the combination is not yet
>>>trustworthy.
>>>      
>>>
>>well - tough one, it works, but the big issue is that you rredologfile
>>need to be contiguous on disk, otherwise you might have 
>>failures, exact
>>same goes for rhel3 as rhas21. you can see that by running debugocfs
>>eg :
>>/ocfs/log/foo1.dbf -> debugocfs -f /log/foo1.dbf /dev/sdXXX
>>that will show how many offsets (should only have one) in the extents
>>if its more than 1, dd it over with a very large blocksize and see if
>>that ends up being 1 contig file.
>>
>>if you do that, everything should work, however, there just 
>>hasn't been
>>enough real testing with aio, need to ggather more evidence.
>>
>>the reason the logfiles are annoying is because he way aio is
>>implemented and how we call it, it cannto handle short io's or non
>>contig aio submits.
>>    
>>
>_______________________________________________
>Ocfs-users mailing list
>Ocfs-users at oss.oracle.com
>http://oss.oracle.com/mailman/listinfo/ocfs-users
>  
>