[nSLUG] hard disks can't be trusted (as if you didn't already know that)
nslug at kernelpanic.ca
Sun Mar 4 18:19:05 AST 2007
On 4-Mar-07, at 16:27 , George N. White III wrote:
> Google's study is for consumer grade disks, but their overall
> failure rates aren't out of line with other studies.
The Google study is interest, but there is a slight problem: they
don't distinguish between failure types. The study study states that
failures such as controller failure were also counted as drive
failures. This does not make the study worthless, but the conclusions
should be more along the lines of: "more study needed." Probably the
most interesting aspect of the study is that SMART is a poor
indicator of drive health.
A Carnegie Mellon U study published at the same conference is also
interesting. They also found no bath-tub effect. They found that
failure rates were significantly higher then manufacturers claims.
The study found no correlation between failure and SCSI, FCAL, or
SATA. However it is not clear if the SATA drives are desktop drives
or server grade "near-line" drives.
I definitely want to see more studies on consumer grade versus server
grade drives. In the lab we have two Sun raid arrays and we had an
Apple raid. Sun uses stupid expensive SCSI drives. Apple uses cheap
desktop (not even semi-server) IDE drives in FCAL to IDE sleds. The
drives in the Apple raid failed so often that it was decommissioned
in less then a year. There are twice as many SUN drives in operation
since 2001 and I think only one drive has ever failed. Obviously
there are not enough drives to draw any conclusions from, but it has
made us reconsider using cheaper desktop hard drives. (It doesn't
hurt that SUN, IBM, Dell, and HP will courier a drive to me in 1 day
and Apple won't.)
More information about the nSLUG