On 18/04/12 13:39, David Brown wrote:
On 18/04/2012 13:52, Benjamin ESTRABAUD wrote:
Hi,
I was wondering about the following:
Superblocks, and all RAID metadata, are stored on disks (to assemble the
RAID), and also on the RAID (while assembled), and are necessary to run
a RAID correctly, so long as at least <parity reached> of superblocks on
disks are available, as <parity reached> number of disks are required
for a specific RAID level to run (this excludes RAID 0 obviously).
This means that so long as less than 1 disk fails in RAID5, no more than
one superblock will be lost and therefore the RAID can still assemble,
and the metadata be read.
However, in modern RAID systems, the disks are all connected through a
single path, being a SAS cable connected to a JBOD or a single SATA
controller that can fail/crash.
Also, the RAID is not protected against power failure, which in my head
are a bit equivalent to a complete disk link failure (SAS cable pulled).
In these cases where all the disks are lost at once, what is the
probability of superblock corruption (both on the RAID superblock and
the individual disks)?
If the superblock was being written during the failure, would it be
incompletely written and therefore corrupted?
How reliably is it to keep a RAID alive (being able to re-assemble it)
after continuously pulling and pushing the SAS cable?
Regards,
Ben.
I think the idea of RAID is that it is a "redundant array of
inexpensive disks" - it is the disks that are redundant. If you get
sudden failure of other parts of the system, there is always a risk of
corrupting or losing the whole array. So it is a question of what are
the likely failures, and how is it best to guard against them -
minimising the chances of failure and the consequences of failure.
In general, the Linux md raid, the block drivers, the filesystems,
etc., are as robust as reasonably practical in the face of unexpected
power failures or other crashes. There are also some filesystems that
can be balanced here (such as with xfs, choosing to enable or disable
write barriers) between speed and safety. But I don't think anyone
gives you concrete guarantees.
In real life, there are three things that are likely failures. Hard
disks (and SSD's) can fail - RAID protects you here. The power supply
on the server can fail - if you worry about that, most quality servers
can be fitted with redundant power supplies. And the external power
supply can fail - for that, you use an UPS.
I have never heard of a SAS or SATA cable failing - I think you would
have to abuse it vigorously to cause damage. Controller cards /can/
fail - the more complex they are, the bigger the risk. I have never
seen a SATA controller fail, but I did once have a SAS card that
failed. If you see this sort of thing as a risk, then make sure you
have two controllers, and your array is built using mirrors at some
level, with each half of the mirror(s) on separate controllers.
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi David,
Thanks for your reply.
I'm actually trying here to determine some probabilities and risk
related to RAID and superblocks.
I understand that the failures are unlikely and solutions can be
implemented to prevent failures, but I'm trying to figure out the risks
involved in various scenarios, like SAS cable failure (pulled out by a
human), controller failure (can happen, even if rare) etc.
Filesystem here is not an issue either, this is really about MD and its
sturdiness to various failures.
Considering the metadatas are stored "inband" (on the same device as the
RAID data), what are the chances of breaking an entire RAID's metadata
because of a massive failure (all disks failing or SAS/SATA controller
failing or power failure)? I understand that the last MD BIOs may be
lost, that some filesystem writes may be lost or corrupted, but I'm more
interested about the superblock corruption here that would cause the
RAID never to re-assemble without a drastic action like an entire
re-create with or without --clean.
Would you know more about that?
Regards,
Ben.
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
