On Wed, 2011-02-16 at 15:56 -0500, Greg Smith wrote:
> Bryan Keller wrote:
> > It sounds like NFS is a viable solution nowadays. I a still going to shoot for using iSCSI, given it is a
block-levelprotocol rather than file-level, it seems to me it would be better suited to database I/O.
> >
>
> Please digest carefully where Joe Conway pointed out that it took them
> major kernel-level work to get NFS working reliably on Linux. On
> anything but Solaris, I consider NFS a major risk still; nothing has
> improved "nowadays" relative to when people used to report regular
> database corruption running it on other operating systems. Make sure
> you read
> http://www.time-travellers.org/shane/papers/NFS_considered_harmful.html
> and mull over the warnings in there before you assume it will work, too.
>
> I don't think I've ever heard from someone happy with an iSCSI
> deployment, either. The only way you could make an NFS+iSCSI storage
> solution worse is to also use RAID5 on the NAS.
>
> I'd suggest taking a look at
> http://wiki.postgresql.org/wiki/Shared_Storage and consider how you're
> going to handle fencing issues as well here. One of the reasons SANs
> tend to be preferred in this area is because fencing at the
> fiber-channel switch level is pretty straightforward. DAS running over
> fiber-channel can offer the same basic features though, it's just not as
> common to use a switch in that environment.
In short, use DAS or a SAN. iSCSI suffers from all kinds of performance
issues and NFS is just Michael Myers scary.
With DAS systems able to handle up 192 drives over 6Gb/s a second these
days, combined with a volume manager you can solve a lot of problems
without breaking the bank.
JD
--
PostgreSQL.org Major Contributor
Command Prompt, Inc: http://www.commandprompt.com/ - 509.416.6579
Consulting, Training, Support, Custom Development, Engineering
http://twitter.com/cmdpromptinc | http://identi.ca/commandprompt