CLVM + DRBD dual primary? (was Re: replicated file system?)

[David Miller]

On Wed, Feb 29, 2012 at 11:14 AM, Alan Johnson <alan at datdec.com> wrote:

> On Wed, Feb 29, 2012 at 10:32 AM, Mark Komarinski <mkomarinski at wayga.org>wrote:
>
>> As pointed out before, DRBD can do active/active, so long as the
>> filesystem on top support it (GFS2 and OCFS2).  The DRBD team even has
>> documentation to get you started:
>>
>> http://www.drbd.org/users-guide/ch-gfs.html
>> http://www.drbd.org/users-guide/ch-ocfs2.html
>>
>
> The "active/active" terminology just clicked for me.  (brain fart! I was
> pretty tired when I was responding before.)  While I am very intrigued by
> this dual-primary functionality in DRBD, I expect gluster will work out
> bitter-faster-stronger for Kenny in my experience, even if it would not
> bring you a bed pan and breakfast as beens seems to have interpreted my
> previous comments to include.   ;-p  [Ass-ide: Of course, Ben is right.
> Only certain drugs can make "everything trivial", and they don't last.]  =)
>
> That said, does anyone know if CLVM would meet the needs of a dual-primary
> DRBD?  I'm having trouble finding good docs on clustering LVM.  Maybe I
> just need to look again now that I'm more awake! =)  Still, any hints would
> be appreciated.
>
> If it matters, the LVs would only be accessed by one node at a time.  They
> would serve as backing block devices for KVM virtual drives enabling live
> migration.  This could remove the need for iSCSI and heartbeat from my
> current DRBD>LVM>iSCSI>heartbeat setup.
>
> If this works as expected, it would be ideal for a 2 host cloud, but live
> migration would not be possible outside those 2 hosts.  Unless... can
> DRBD+CLVM be multi-primary?  Of course, this quickly leads to excessive
> redundancy... unless... DRBD can do something like RAID5/6/Z/Z2, but I'm
> sure I'm asking too much there, especially if I want it to be efficient
> over a network.
>
> I can imagine that CLVM across a bunch of iSCSI PVs would be ideal,
> allowing mirroring and/or stripping on a per LV basis, but I again, I can't
> find good docs on CLVM to back me up.
>
>
To my knowledge DBRD can only do mirroring.  So no higher RAID levels are
supported.

CLVMs role is to deal with locking and syncing of the LVM metadata on VGs
and block devices (PVs) that are visible to multiple hosts.  So if you're
planning on using LVM on top of DBRD block devices then CLVM would probably
be a good thing to use.  But I have heard in the pass that you're careful
about only managing the LVM from one host you may be able to get away with
using the standard LVM tools.  But this seems risky to me and I have never
put that to the test so if you're planning on using that in production you
should verify and test that.

The resulting block devices (LVs) would still need to either be only
mounted by one host or use a clustered file system to provide proper a DLM.
 So I think this would work for Virtual Machines.  You could gain the
extra redundancy by running RAID under DBRD.  RAID (Hardware or Software) -
DBRD - CLVM.  I would also run some sort of network link aggregation
between those two hosts for speed and redundancy at the network layer.
--
David
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.gnhlug.org/mailman/private/gnhlug-discuss/attachments/20120229/28469c92/attachment.html