[PLUG] XFS experiences

[Eli Stair]

BTW, if you're looking almost exclusively for lazy-write/short-lived
small file optimizations, I've hacked that up with ext3 succesfully. 
Once you have a large & fast external journal device on disks (raid10)
_separate_ from your main filesystem array, you use that for all your
filesystem writes first (data=journal).  Inline with that, setting a
very large journal flush period enables short-lived files to be kept
off of your main array... and all that IO directed at the disk doesn't
affect read performance until it does a large batch write of long-term
data (MUCH better than sporadic small writes).

/eli

On 4/20/06, Roderick A. Anderson <raanders at acm.org> wrote:
> Thanks Eli.  Very interesting.  The warm fuzzy I was reading came from
> an IBM site.  Nothing about JFS though.  :-)  I guess I was impressed
> with the 8-9 EByte size and the lazy write feature.
>
> So minus-one for XFS.  Plus-two for ext3.
>
> Anyone else?
>
>
> Rod
> --
> Eli Stair wrote:
> > I'll give a brief brief, if you want more info I'll be happy to
> > provide details ;)
> >
> > XFS is great.  However, so is ext3.  (very non-commital ;)
> >
> > In the event that everything works OK, I've been extremely happy with
> > XFS.  When I've had failures they are tremendously huge, and sometimes
> > have a creeping-doom leaving up to them making recovery painful.  I
> > used to rant about how inferior ext3 was, and scoff at those to didn't
> > take the time to test and tweek _every_ filesystem tunable and
> > generate pretty graphs documenting superiority... I'm recovering from
> > that now!
> >
> > After unaddressed significant XFS bugs lingering through numerous
> > in-kernel releases and XFS project releases, I realized it wasn't
> > something that I wanted to be running on critical systems.  I've had
> > issues with kernel bugs on ext3, but they've been addressed _swiftly_
> > by Redhat and the developers.  XFS had numerous bugzilla entries from
> > me for a long time, that never responded...
> >
> > Pros:
> >
> > XFS:
> >   Fast.  Tunable.  Chic.
> >
> > ext3:
> >   Fast.
> >   Extremely well-tested and proven recovery/lint tools.
> >   Great full-time coder/vendor support.
> >
> > Cons:
> >
> > XFS:
> >   If it breaks, plan on recovering from backup and switching to a
> > different filesystem.  If it's a repeatable bug, you'll hit it
> > again... and again... and post how repeatable it is... and get no
> > response.
> >
> > ext3:
> >   The docs are only like 10% as big ;)
> >   Far fewer filesystem layout options.
> >
> > Breakdown:
> >   ext3 is fast when set up right, and very stable.  If you've got the
> > new hardware investment, there are a number of things you can do to
> > tweak your layout properly for your workload.   If you'd like to chat
> > about it, let me know.  If you want to play with a system specifically
> > for non-critical temporary data with massively concurrent IO (and you
> > have a LOT of money to make that work), go with XFS (or just buy some
> > Spinnakers).  If you want a really capable, stable, and warm-fuzzy
> > system that is capable of massively concurrent IO (within the limits
> > of your hardware), I'd go with ext3 (or just buy some Spinnakers)  and
> > put a bit of effort (a couple days) into planning and testing it.
> >
> > IMO.
> >
> > Cheers,
> >
> > /eli
> >
> >
> >
> >
> > On 4/20/06, Roderick A. Anderson <raanders at acm.org> wrote:
> >
> >>Ref: SCSI vs. SATA RAID messages.
> >>
> >>I'm down to the point of ordering our new ( specialized ) server so I'm
> >>thinking of other performance considerations for the software it will be
> >>running.  This brings up file systems.
> >>    I typically just choose ext3 and SWAG a partition scheme depending
> >>on how much disk space the system has.  This has worked for well for the
> >>general purpose systems I've build to date.  Since the software we'll be
> >>running is very disk and RAM intensive we've loaded up ( given the
> >>projected work load ) on RAM and the SCSI RAID will help but I'm now
> >>looking at file systems.
> >>    Choices fall as either ext3 and XFS.  I know how ext3 works and it's
> >>warts but I've never used XFS.  The reason I'm not considering ReiserFS
> >>  is we have million's of files ranging in size from a few KByte to a
> >>few GByte in size.
> >>    There are also a lot of temporary files that are short lived so --
> >>according to the XFS articles I've read -- will never hit the disks.
> >>
> >>What are your thoughts and ideas on XFS as a file system.
> >>
> >>
> >>TIA,
> >>Rod
> >>--
> >>_______________________________________________
> >>PLUG mailing list
> >>PLUG at lists.pdxlinux.org
> >>http://lists.pdxlinux.org/mailman/listinfo/plug
> >>
> >
> > _______________________________________________
> > PLUG mailing list
> > PLUG at lists.pdxlinux.org
> > http://lists.pdxlinux.org/mailman/listinfo/plug
>
> _______________________________________________
> PLUG mailing list
> PLUG at lists.pdxlinux.org
> http://lists.pdxlinux.org/mailman/listinfo/plug
>