[Sheepdog] Sheepdog+iscsi high availability

[joby xavier]

Thanks for the informations Yuan.

Regards,
Joby Xavier

On Fri, Apr 20, 2012 at 4:11 PM, Liu Yuan <namei.unix at gmail.com> wrote:

> On 04/20/2012 05:56 PM, joby xavier wrote:
>
> > Hi Yuan,
> >
> > I  tested sheepdog with the following steps
> > Total sheep nodes : 5
> >
> >  mount /dev/sda3 /var/lib/sheepdog/
> >  mount -o remount,user_xattr /var/lib/sheepdog
> >  sheep -d /var/lib/sheepdog
> >
> >  collie cluster format -b farm (only on one node)
> >
> > Then i create sheep volume for iscsi
> >
> >  qemu-img create sheepdog:tom -o preallocation=data 2G
> >
> >  collie vdi list
> >   Name        Id    Size    Used  Shared    Creation time   VDI id  Tag
> >   tom          1  2.0 GB  2.0 GB  0.0 MB 2012-04-20 14:06   65958b
> >
> > collie node info
> > Id    Size    Used    Use%
> >  0    20 GB    400 MB      2%
> >  1    20 GB    488 MB      2%
> >  2    20 GB    300 MB      1%
> >  3    20 GB    360 MB      1%
> >  4    20 GB    512 MB      2%
> > Total    98 GB    2.0 GB      2%
> >
> >
> > then sharing sheep volumes on two nodes through iscsi tgt
> >
> > tgtadm --op new --mode logicalunit --tid 1 --lun 1 -b tom --bstype
> sheepdog
> >
> > then, initiator server access iscsi through virtual IP
> > starts VM on this iscsi+sheep volume
> >
> > What i see is that even one node fails the os gets crash and gets medium
> > errors. I tried both cache=none, writeback
> >
>
>
> Frankly speaking, I have no idea to use sheepdog with iscsi protocol and
> never have tried iscsi setup.
>
> I can only figure out that you separate storage (Sheepdog) from
> computing (VM), and in your computing node, there is no any sheep deamon
> for routing requests. Because of absence of sheep daemon, you can't take
> advantage of object cache, which acts as a local cache that supposed to
> reduce the network traffic and data availability.
>
> Without sheep daemon on the node, 'cache' option for qemu actually
> doesn't make any difference.
>
> Currently, even if you want to run VM in a node that doesn't want to
> provide storage to cluster, you can start sheep daemon with
>
> sheep -v 0
>
> on the dedicated node, which means you just use sheep as a gateway for
> your VM. With this scheme, you can make good use of object cache, live
> migration, snapshot, clone features.
>
> > would you please throw some light on this? please let me know if you
> > need any more logs or details. When can we expect the stable version of
> > sheepdog?
>
>
> In our usage(node both provide storage and computing), sheepdog acts
> quit good in the node fail/join events, with this context, I could say
> sheepdog is very close to stable (we need a long running cluster to
> prove this).
>
> > We would like to roll this out to our production servers for high
> > availability VMs and storage.
>
>
> There are someone asking to implement sheep client driver into kernel
> instead of qemu block layer, exporting a block volume in /dev/sheep_volume.
>
> We are happy to get patches to implement this feature, but currently we
> don't have hands for it.
>
> I am considering to implement a sheepfs based on FUSE, to both export
> sheep internal information (what collie does for now) and export image
> block volume (similar to /dev/sheep_volume). But I can't promise when I
> can start/finish it, because I am more concerned of scalability. (we
> expect to run a cluster more than 1000 nodes)
>
> Thanks,
> Yuan
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.wpkg.org/pipermail/sheepdog/attachments/20120420/c7758646/attachment-0003.html>