[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Questions about RAID 6

Disclaimer:  I'm partial to XFS

Tim Clewlow put forth on 5/1/2010 2:44 AM:

> My reticence to use ext4 / xfs has been due to long cache before
> write times being claimed as dangerous in the event of kernel lockup
> / power outage. 

This is a problem with the Linux buffer cache implementation, not any one
filesystem.  The problem isn't the code itself, but the fact it is a trade
off between performance and data integrity.  No journaling filesystem will
prevent the loss of data in the Linux buffer cache when the machine crashes.
 What they will do is zero out or delete any files that were not fully
written before the crash in order to keep the FS in a consistent state.  You
will always lose data that's in flight, but your FS won't get corrupted due
to the journal replay after reboot.  If you are seriously concerned about
loss of write data that is in the buffer cache when the system crashes, you
should mount your filesystems with "-o sync" in the fstab options so all
writes get flushed to disk without being queued in the buffer cache.

> There are also reports (albeit perhaps somewhat
> dated) that ext4/xfs still have a few small but important bugs to be
> ironed out - I'd be very happy to hear if people have experience
> demonstrating this is no longer true. My preference would be ext4
> instead of xfs as I believe (just my opinion) this is most likely to
> become the successor to ext3 in the future.

I can't speak well to EXT4, but XFS has been fully production quality for
many years, since 1993 on Irix when it was introduced, and since ~2001 on
Linux.  There was a bug identified that resulted in fs inconsistency after a
crash which was fixed in 2007.  All bug fix work since has dealt with minor
issues unrelated to data integrity.  Most of the code fix work for quite
some time now has been cleanup work, optimizations, and writing better
documentation.  Reading the posts to the XFS mailing list is very
informative as to the quality and performance of the code.  XFS has some
really sharp devs.  Most are current or former SGI engineers.

> I have been wanting to know if ext3 can handle >16TB fs.  I now know
> that delayed allocation / writes can be turned off in ext4 (among
> other tuning options I'm looking at), and with ext4, fs sizes are no
> longer a question. So I'm really hoping that ext4 is the way I can
> go.

XFS has even more tuning options than EXT4--pretty much every FS for that
matter.  With XFS on a 32 bit kernel the max FS and file size is 16TB.  On a
64 bit kernel it is 9 exabytes each.  XFS is a better solution than EXT4 at
this point.  Ted T'so admits last week that one function call in EXT4 is in
terrible shape and will a lot of work to fix:

"On my todo list is to fix ext4 to not call write_cache_pages() at all.
We are seriously abusing that function ATM, since we're not actually
writing the pages when we call write_cache_pages().  I won't go into
what we're doing, because it's too embarassing, but suffice it to say
that we end up calling pagevec_lookup() or pagevec_lookup_tag()
*four*, count them *four* times while trying to do writeback.

I have a simple patch that gives ext4 our own copy of
write_cache_pages(), and then simplifies it a lot, and fixes a bunch
of problems, but then I discarded it in favor of fundamentally redoing
how we do writeback at all, but it's going to take a while to get
things completely right.  But I am working to try to fix this."

> I'm also hoping that a cpu/motherboard with suitable grunt and fsb
> bandwidth could reduce performance problems with software raid6. If
> I'm seriously mistaken then I'd love to know beforehand. My
> reticence to use hw raid is that it seems like adding one more point
> of possible failure, but I could be easily be paranoid in dismissing
> it for that reason.

Good hardware RAID cards are really nice and give you some features you
can't really get with md raid such as true "just yank the drive tray out"
hot swap capability.  I've not tried it, but I've read that md raid doesn't
like it when you just yank an active drive.  Fault LED drive, audible
warnings, are also nice with HW RAID solutions.  The other main advantage is
performance.  Decent HW RAID is almost always faster than md raid, sometimes
by a factor of 5 or more depending on the disk count and RAID level.
Typically good HW RAID really trounces md raid performance at levels such as
5, 6, 50, 60, basically anything requiring parity calculations.

Sounds like you're more of a casual user who needs lots of protected disk
space but not necessarily absolute blazing speed.  Linux RAID should be fine.

Take a closer look at XFS before making your decision on a FS for this
array.  It's got a whole lot to like, and it has features to exactly tune
XFS to your mdadm RAID setup.  In fact it's usually automatically done for
you as mkfs.xfs queries the block device device driver for stride and width
info, then matches it.  (~$ man 8 mkfs.xfs)

(note the date, and note the praise Hans Reiser lavishes upon XFS)
(2.6.34-rc3 benchmarks, all filesystems in tree)

XFS Users:

 The Linux Kernel Archives

"A bit more than a year ago (as of October 2008) kernel.org, in an ever
increasing need to squeeze more performance out of it's machines, made the
leap of migrating the primary mirror machines (mirrors.kernel.org) to XFS.
We site a number of reasons including fscking 5.5T of disk is long and
painful, we were hitting various cache issues, and we were seeking better
performance out of our file system."

"After initial tests looked positive we made the jump, and have been quite
happy with the results. With an instant increase in performance and
throughput, as well as the worst xfs_check we've ever seen taking 10
minutes, we were quite happy. Subsequently we've moved all primary mirroring
file-systems to XFS, including www.kernel.org , and mirrors.kernel.org. With
an average constant movement of about 400mbps around the world, and with
peaks into the 3.1gbps range serving thousands of users simultaneously it's
been a file system that has taken the brunt we can throw at it and held up


Reply to: