Re: Database storage - Mailing list pgsql-general

In response to nabble.30.miller_2555@spamgourmet.com:

> > On Thu, Jul 9, 2009 at 7:29 PM, Greg Stark<gsstark@mit.edu> wrote:
> >> On Fri, Jul 10, 2009 at 1:28 AM, Scott Marlowe<scott.marlowe@gmail.com> wrote:
> >>>
> >>> $750 is about what a decent RAID controller would cost you, but again
> >>> it's likely that given your bulk import scenario,  you're probably ok
> >>> without one.  In this instance, you're probably best off with software
> >>> RAID than a cheap RAID card which will cost extra and probably be
> >>> slower than linux software RAID.
> ...
> >> The main advantage of hardware raid is the error handling. When you
> >> get low level errors or pull a drive a lot of consumer level
> >> controllers and their drivers don't respond very well and have long
> >> timeouts or keep retrying tragically unaware that the software raid
> >> would be able to handle recoverying. A good server-class RAID
> >> controller should handle those situations without breaking a sweat.
>
> > Definitely a big plus of a quality HW controller, and one of the
> > reasons I don't scrimp on the HW controllers I put in our 24/7
> > servers.  OTOH, if you can afford a bit of downtime to handle
> > failures, linux software RAID works pretty well, and since quad core
> > CPUs are now pretty much the standard, it's ok if parity calculation
> > uses up a bit of one core for lower performing servers like the
> > reporting server the OP was talking about.
>
> The database server is a quad core machine, so it sounds as though
> software RAID should work fine for the present setup. However, it
> sounds as though I should put some money into a hardware RAID
> controller if the database becomes more active. I had assumed RAID-5
> would be fine, but please let me know if there is another RAID level
> more appropriate for this implementation. Thanks for the valuable
> insight!

RAID 10 is pretty much the fastest RAID level for disk IO.  Every pair
of disks you add to a RAID-10 array makes the array faster (assuming
you don't hit any controller bottlenecks)

Another advantage of hardware RAID controllers (that I'm surprised nobody
has mentioned) is battery-backed cache.  Using said cache, you can
configure the controller to lie about fsyncs, which make them essentially
free from PostgreSQL's standpoint.  Since the cache is backed by a
battery, your concerns about data loss in the event of power failure are
much less.  The cache doesn't usually increase the overall throughput of
the system, but it usually improves peak load performance by deferring
writes until things are calmer.

--
Bill Moran
http://www.potentialtech.com
http://people.collaborativefusion.com/~wmoran/