On Jan 31, 2006 11:45 -0600, Pappas, Bill wrote:> Has anyone considered using luster as the foundation to export nfs and
> samba shares?
Yes, this is possible to do already, with some caveats.
- NFS exporting of 2.4 lustre clients works (though not as fast as Lustre)
- NFS exporting of 2.6 lustre clients works with user-space NFSD
- samba exporting of Lustre works
> Can anyone explain how osts can be set up to protect data from ost and
> storage target failure?
>
> Assume you have a fiber channel san. I''ve presented four disks
from one
> storage array to two osts, but under normal conditions each ost uses 2
> of the four luns. I''''ve created a OBD that is a four way
stripe where
> each stripe uses one (of the four) LUN from the storage array. Is this
> possible where two OSTs can work together to comprise a OBD stripe? Is
> this called an active-active OST setup?
No - Lustre OSTs (and MDSes) should NEVER, EVER access the same LUN at
the same time. That is a surefire way to data corruption. Each OST
only "exports" a single filesystem/device at a time. An active-active
OSS pair means that each OSS node is exporting "some" OST, but they
never export the same OST at the same time.
> OST Failure: If one ost fails (and times out back to the client) will
> the other ost pick up the two luns (since they were zoned/masked via the
> fabric to be seen in the first place) and maintain this 4-way stripe on
> itself?
This is currently handled external to Lustre, via a package like
Heartbeat or CluManager. In order to ensure only a single OSS is
accessing the shared storage at one time you need to have power
control for the nodes, so the "takeover" OSS can STONITH the
"failing"
OSS before accessing its devices.
> Storage Target Faulure: Can I add two more OSTs, isolated from the
> other two OSTs and set them up as a mirror for the previous two OSTs and
> its respective ODB stripe. I''m trying to create a mirrored-stripe
using
> lustrue across two separate storage arrays. I realize there is an IO
> penalty for mirroring as two writes would have to be committed. I
> wonder if the read IO would increase as more OSTs are delivering the
> ODB? With hardware and software mirroring, there can be a boost in read
> IO. Am I even on the same page? Basically, I am looking to see if
> luster can offer protection against storage susbsystem (target) failure.
Lustre is not currently able to do RAID1 striping itself (in the LOV layer).
It will only do RAID0 striping. You are strongly recommended to have some
sort of redundancy at the device level (hardware/software RAID1 or RAID5).
Cheers, Andreas
--
Andreas Dilger
Principal Software Engineer
Cluster File Systems, Inc.