Re: Solid State Drives
Date: Fri, 01 May 2009 10:27:42 -0500
We're in the middle of discussions about this too (just had another one yesterday), and we have started deploying PCIe-based local devices for a few applications. We don't have an internal consensus yet about mirroring. Our Engineering group (my group) is currently recommending to mirror the devices, but I think a few business customers have chosen not to. Seems that mitigating the risk of data loss isn't always worth the cost to them.
We are managing them with ASM, so no worries about filesystem block sizes. And of course, as Matt says, the enterprise-level SSDs do write-leveling at the hardware level. I'm not sure if anyone's mentioned yet that there are also power savings with SSD - especially over the rack of disks needed to get comparable IOPS. But it's still worth pointing out that right now (as far as I can tell), flash-based SSD is not significantly more performant than HDD for large sequential write operations (e.g. redo/archive), especially when taking price into consideration.
For me, the bigger question is how to best utilize them with currently shipping Oracle databases. For small databases, you could possible stick the whole thing on there. (But is it worth the cost?) For larger databases, it's probably application dependent... redo? temp? undo? hot datafiles? Unfortunately, my job right now is to try and figure out best practices and general recommendations which will be read by lots of application groups as they decide whether or not to use this hardware. And most of them don't know much about how Oracle works internally... so we have to provide solid yet basic/general guidance. It's a terrible predicament really. :)
Matthew Zito wrote:
> Bear in mind that enterprise SSDs are already write-leveled at a
> hardware level. The big optimization for filesystems with enterprise
> SSD is making sure that filesystem blocks are aligned with the same
> block structure as the SSD is using – for cheap SSDs that don’t do
> write leveling, then there’s a lot of CoW optimizations you can do.
> And you would still RAID your disks – write leveling is nice, but it
> doesn’t handle the failure of a controller chip or something similar.
> So you’re never going to get away from that. As far as the upgrade
> path, the lifespan is comparable for a “spinning rust” hard drive.
-- Jeremy Schneider Chicago, IL http://www.ardentperf.com -- http://www.freelists.org/webpage/oracle-lReceived on Fri May 01 2009 - 10:27:42 CDT