Re: [Gluster-devel] Question

[Anand Babu Periasamy]

> Hello,
>
> I was reading the Glusterfs and it's really a nice concept for
> shared storage but is it possible to inverse the layers? What I mean
> is that Glusterfs is on top of a ext3/xfs fs, how about the oposite?

To do reverse, GlusterFS has to provide a raw block level access, by
just aggregating the underlying block devices. Even though it is
easier to achieve, benefits are not attractive. 

* RAID-0/1/10:
 - Scaling in capacity leads to painfully longer fsck downtime and
   possibilities for corruption.
 - Scaling in I/O bandwidth is bottle-necked by client's NIC. Because
   block-devices cannot be shared, a single client cannot take
   advantage of aggregated I/O bandwidth.
* RAID-3/4/5/6/XXX 
 - Inherits the above problems.
 - Checksum calculation will consume serious CPU cycles at the client
   side and cause slow down.

They are the same scaling problems SAN technology is facing
currently. To build a scalable powerful file system, clustering has to
be done at file I/O level.

> I was thinking about it because I don't know how Glusterfs reacts if
> it was built on top of a MD+LVM+XFS and then with resizing and
> snapshotting behind Glusterfs back.

MD+LVM+XFS should be transparent to GlusterFS. I don't see any issues
in doing so.

Actually, instead of MD/LVM, GlusterFS..
* can cluster across systems.
* can scale on demand both in terms of capacity and I/O performance
  by adding more storage bricks(servers).
* can snapshot distributedly (coming in 1.4). GlusterFS's snapshot
  will allow you even roll back selected files. 

--
Anand Babu
GPG Key ID: 0x62E15A31
Blog [http://ab.freeshell.org]
The GNU Operating System [http://www.gnu.org]