RE: Datafiles on nfs mounts.

>>We've been running databases on NFS mounts on a Netapp for a number of
>>years. Very reliable. Not the fastest, but not bad.

...thanks Steve, but having a few folks trickle in with testimonals will make it seem like Oracle over NFS is a "sometimes man bites dog" situation. The fact is that there are a tremendous number of Oracle over NFS deployments perhaps most notably Oracle On Demand. Yes, that's right. If you shell out money for Oracle to host your database, they stuff it in a filer.

That leads me to the second part of your posts, speed. Oracle over NFS performance is affected by several things. First, the architecture. It doesn't take much to saturate a filer--after all the insides of current filer technology is a simple x86 machine (some are Xeon, some are AMD). So if you are hammering a filer from another system of the same architecture, there is a sort of parity between system and storage bandwidth right out of the gate. Worse, if you put a RAC database on a single filer and hammer on it with 2 or more modern x86 boxes, you have an obvious bottleneck. This is why clustered storage is on the rise--to offer active:active read/write access to the same directories and FILES in an NFS export. The other thing that will bottleneck an Oracle deployment on filers is the NVRAM cache. So folks often buyu another filer just for logging...then you start to get into "filer sprawl"..yet another need for clsutered storage.

I have an entire section on my blog about this topic and no, it isn't sales material :-)
http://kevinclosson.wordpress.com/kevin-closson-index/cfs-nfs-asm-topics /

>>rw,hard,intr,proto=tcp,vers=3,rsize=32768,wsize=32768,forcedirectio

>>There is discussion if the 'forcedirectio' is needed if you set the
filesystemio_options=setall.

If Solaris pushed directio(3C) onto anything other than RAW, it uses library-threaded async IO. I'm not knocking library-threaded AIO, but it is a bit more expensive than KAIO. In the end, if Oracle successfully calls directio(3C) on an NFS mount, then you get "OS" AIO which is better than multiple DBWR or slaves.

If someone could run the probram listed in http://kevinclosson.wordpress.com/2007/01/11/analysis-and-workaround-for -the-solaris-10203-patchset-problem-on-vxfs-files/ on a file in a Solaris NFS mount (mounted with forcedirectio option), we'd know for certain whether the OS supports it. Once we know that, then we know Oracle *could* do AIO on NFS files if it wanted to--but that would be up to the OSDs.

--
http://www.freelists.org/webpage/oracle-l