• ZFS
    November 18, 2005


Guest Author

The original promise of RAID (Redundant Arrays of Inexpensive Disks)
was that it would provide fast, reliable storage using cheap disks.
The key point was cheap; yet somehow we ended up
here. Why?

RAID-5 (and other data/parity schemes such as RAID-4, RAID-6, even-odd,
and Row Diagonal Parity) never quite delivered on the RAID promise -- and can't --
due to a fatal flaw known as the RAID-5 write hole. Whenever you update the
data in a RAID stripe you must also update the parity, so that all disks XOR
to zero -- it's that equation that allows you to reconstruct data when a
disk fails. The problem is that there's no way to update two or more disks
atomically, so RAID stripes can become damaged during a crash or power outage.

To see this, suppose you lose power after writing a data block but before
writing the corresponding parity block. Now the data and parity for that
stripe are inconsistent, and they'll remain inconsistent forever (unless
you happen to overwrite the old data with a full-stripe write at some point).
Therefore, if a disk fails, the RAID reconstruction process will generate
garbage the next time you read any block on that stripe. What's worse,
it will do so silently -- it has no idea that it's giving you corrupt data.

There are software-only workarounds for this, but they're so slow that
software RAID has died in the marketplace. Current RAID products all do
the RAID logic in hardware, where they can use NVRAM to survive power loss.
This works, but it's expensive.

There's also a nasty performance problem with existing RAID schemes.
When you do a partial-stripe write -- that is, when you update less data
than a single RAID stripe contains -- the RAID system must read the old data
and parity in order to compute the new parity. That's a huge performance hit.
Where a full-stripe write can simply issue all the writes asynchronously,
a partial-stripe write must do synchronous reads before it can even start
the writes.

Once again, expensive hardware offers a solution: a RAID array
can buffer partial-stripe writes in NVRAM while it's waiting for the
disk reads to complete, so the read latency is hidden from the user.
Of course, this only works until the NVRAM buffer fills up. No problem,
your storage vendor says! Just shell out even more cash for more NVRAM.
There's no problem your wallet can't solve.

Partial-stripe writes pose an additional problem for a transactional
filesystem like ZFS. A partial-stripe write necessarily modifies live data,
which violates one of the rules that ensures transactional semantics.
(It doesn't matter if you lose power during a full-stripe
write for the same reason that it doesn't matter if you lose power during
any other write in ZFS: none of the blocks you're writing to are live yet.)

If only we didn't have to do those evil partial-stripe writes...

Enter RAID-Z.

RAID-Z is a data/parity scheme like RAID-5, but it uses dynamic stripe width.
Every block is its own RAID-Z stripe, regardless of blocksize. This means
that every RAID-Z write is a full-stripe write. This, when combined with the
copy-on-write transactional semantics of ZFS, completely eliminates the
RAID write hole.
RAID-Z is also faster than traditional RAID because it never has to do

Whoa, whoa, whoa -- that's it? Variable stripe width? Geez, that seems
pretty obvious. If it's such a good idea, why doesn't everybody do it?

Well, the tricky bit here is RAID-Z reconstruction. Because the stripes
are all different sizes, there's no simple formula like "all the disks
XOR to zero." You have to traverse the filesystem metadata to determine
the RAID-Z geometry. Note that this would be impossible if the filesystem
and the RAID array were separate products, which is why there's nothing
like RAID-Z in the storage market today. You really need an integrated
view of the logical and physical structure of the data to pull it off.

But wait, you say: isn't that slow? Isn't it expensive to traverse
all the metadata? Actually, it's a trade-off. If your storage pool
is very close to full, then yes, it's slower. But if it's not too
close to full, then metadata-driven reconstruction is actually faster
because it only copies live data; it doesn't waste time copying
unallocated disk space.

But far more important, going through the metadata means that ZFS
can validate every block against its 256-bit checksum as it goes.
Traditional RAID products can't do this; they simply XOR the data
together blindly.

Which brings us to the coolest thing about RAID-Z: self-healing data.
In addition to handling whole-disk failure, RAID-Z can also detect
and correct silent data corruption. Whenever you read a RAID-Z block,
ZFS compares it against its checksum. If the data disks didn't return
the right answer, ZFS reads the parity and then does combinatorial
reconstruction to figure out which disk returned bad data. It then
repairs the damaged disk and returns good data to the application.
ZFS also reports the incident through Solaris FMA so that the system
administrator knows that one of the disks is silently failing.

Finally, note that RAID-Z doesn't require any special hardware.
It doesn't need NVRAM for correctness, and it doesn't need write buffering
for good performance. With RAID-Z, ZFS makes good on the original RAID
promise: it provides fast, reliable storage using cheap, commodity disks.

For a real-world example of RAID-Z detecting and correcting silent
data corruption on flaky hardware, check out
Eric Lowe's SATA saga.

The current RAID-Z algorithm is single-parity, but the RAID-Z concept
works for any RAID flavor. A double-parity version is in the works.

One last thing that fellow programmers will appreciate: the entire
RAID-Z implementation
is just 599 lines.

Technorati Tags:

Join the discussion

Comments ( 7 )
  • Dan Price Friday, November 18, 2005
    I thought you might be gratified to see that RAID-Z is already in wikipedia :)
  • Naveen Nalam Friday, November 18, 2005
    How does ZFS deal with sata drives that have the write cache turned on? Such that in the event of a crash, writes that were acked to the OS are lost?
  • Jeff Bonwick Saturday, November 19, 2005
    We issue the SYNCHRONIZE CACHE command to the disks after pushing all data in a transaction group, but before updating the uberblock to commit it. I'll describe the transaction group commit process in detail in a future post.
  • Al Hopper Monday, November 21, 2005
    Interesting Stuff Jeff. Another point about conventional RAID-5 is that, in dedicated RAID-5 hardware, you quickly saturate the memory bus between the CPU, disk controllers and dedicated XOR hardware. Not only do you saturate it, but you also have bus contention - since the disks will often generate data simultaneously - but that data has to be, essentially, serialized over the single bus. The solution to achieving high performance in this situation is, as you've said, add more hardware and more system complexity (and cost).
    One of my little "tricks", is to determine the system bus characteristics of the RAID-5 hardware. Once you know this, you \*know\* the maximum performance capabilities of the product!
  • Igor Sunday, November 27, 2005
    Very interesting idea. Is any performance data available for RAID-Z?
  • guest Tuesday, December 20, 2005
    FYI, a RAID system with variable sized parity stripes:
  • Veera Swamy Vallabhaneni Friday, April 14, 2006
    Excellent information
Please enter your name.Please provide a valid email address.Please enter a comment.CAPTCHA challenge response provided was incorrect. Please try again.