Lustre discuss - May 2006 - lustre snapshots

Hi Sage,

The snapshots are taken globally, all the OSS''s and MDSs serving a file
system (one MDS only for Lustre 1.X) make the snapshot
"simultaneously".


There is coordination needed to make this happen and all the clients
receive two "messages" (the messages are implemented using our lock
manager):=20

"I will make a snapshot after you confirm X".=20
X here could be one of the following:
  - "do nothing" in which case the client simply suspends processing
  - "you sync" in which case the the reply is sent afte sync is called
  - "your database syncs" in which case the client makes an upcall to
notify databases

This is a rather obvious cluster extension to snapshot mechanisms used
by file systems or LVM''s using a logical volumes.  You could call
"X"
the policy.

The second message to the client follows after the servers make the
snapshot.  The process of making snapshots is almost instantaneous, so
the entire delay is mostly waiting for cache flushes.

The first thing that will work is snapping a whole file system, but
Lustre is usually used for very large file systems, so we are definitely
exploring the option of taking a snapshot only on a subdirectory basis
and similarly to exclude subdirectories from snapshots.  I expect
neither to be all that complicated. =20

I hope this answers your questions.

- Peter -

> -----Original Message-----
> From: lustre-discuss-admin@lists.clusterfs.com=20
> [mailto:lustre-discuss-admin@lists.clusterfs.com] On Behalf=20
> Of sage weil
> Sent: Saturday, May 29, 2004 5:12 AM
> To: lustre-discuss@lists.clusterfs.com
> Subject: [Lustre-discuss] lustre snapshots
>=20
> Hi,
>=20
> I have a couple quick questions about how snapshots work in=20
> Lustre.  I see that it''s mostly implemented in the OSD=20
> layer.. does that mean that snapshots are always created for=20
> the entire file system?  Does the MDS simply simultaneously=20
> instruct all the OSSs to snapshot their object collections?
>=20
> How are client writes synchronized with respect to a=20
> snapshot?  If you were to, for instance, suspend and flush=20
> database tables to disk, take a snapshot, and then resume, do=20
> special measures need to be taken to ensure file system data=20
> if written through to the OSS before the snapshot is taken? =20
> (Actually, I guess write-through is the default policy..)
>=20
> Thanks!
> sage
>=20
>=20
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss@lists.clusterfs.com
> https://lists.clusterfs.com/mailman/listinfo/lustre-discuss
>=20
>=20
>=20

Hi,

I have a couple quick questions about how snapshots work in Lustre.  I see
that it''s mostly implemented in the OSD layer.. does that mean that
snapshots are always created for the entire file system?  Does the MDS
simply simultaneously instruct all the OSSs to snapshot their
object collections?

How are client writes synchronized with respect to a snapshot?  If you
were to, for instance, suspend and flush database tables to disk, take a
snapshot, and then resume, do special measures need to be taken to ensure
file system data if written through to the OSS before the snapshot is
taken?  (Actually, I guess write-through is the default policy..)

Thanks!
sage