[Pc_Support] Re: Using OS software RAID with FRAID organization, and when not to use it (in case I wasn't clear) ...

[Bryan J. Smith]

Jason Boxman wrote:
> Sector remapping?

It's an ubiquitous SCSI concept that has always been applied to true
hardware RAID cards.

SCSI disks don't use all their tracks/sectors.  They reserve several
tracks and sectors for every cylinder as "extras."  When sectors have
to be re-read and that re-reading reaches a threshold (typically just
a few times), the SCSI host adapter will remap that sector to an extra
and mark the original sector bad.  Early SCSI required you to remap in
a manual procedure, whereas latter SCSI-2 does it automagically during
run-time (that's a mega-oversimplification, and I'm a bit ignorant
and/or forgot all of my SCSI-2 protocol functions/basics).

NOTE:  This is _below_ the filesystem.  So from the standpoint of the
OS filesystem itself, no sectors are bad.

Virtually all true hardware RAID cards do the same with their blocking
(multiple sectors grouped together, which is how the card spans
disks), and more enterprise software RAID does as well.  Because
you're going to have sectors go bad on one disk but not exactly in the
equivalent of the other.  The true hardware RAID hides this from you,
including reserving portions of the disk at volume creation.  Hence
why many people new to 3Ware cards complain about it not using the
whole disk and shaving off a few hundred MBs.

I haven't seen a FRAID card yet that does sector remapping.  In a
nutshell, most RAID-1 implementations -- software and hardware --
don't do blocking at all, so they don't offer remapping anyway.  I
know 3Ware cards do remapping, even on RAID-1, but I'm fairly certain
they avoid it if at all possible and then do it at the end of the
volume where they also store meta-data.  I don't know if the
DeviceMapper will mount a 3Ware RAID-1 volume with remapped sectors,
but since the DeviceMapper does mount 3Ware RAID-10 volumes which does
use blocking, it might support it.

I want to say Linux's MD actually does remapping, because it's really
just an extension of spanning.  Although I might be thinking of LVM2
w/DeviceMapper, and one of the advantages of DeviceMapper over MD.  I
need to do some more research, I'm largely ignorant of what is and
isn't supported these days in various software RAID implementations.
But I just know every FRAID I've looked at doesn't, and relies on the
OS' filesystem to mark its sectors/blocks bad.

Bad sectors are far less common than in year's past.  But your chances
of having one are exponential in any striped/mirrored array by the
number of disks -- and then that's a problem.  Especially for FRAID,
which is using hte "raw" disks and the OS is forced to mark all of the
offset sectors bad (even when some are not).