[sheepdog-users] runtime requirements?

Andrew J. Hobbs ajhobbs at desu.edu
Thu Mar 13 20:38:20 CET 2014


It's there, if only because I'm a slave to conformity.  It's not 
actively used.  Those nodes might be the preferred hosts for qemu-nbd 
though, where virtualization support isn't actually needed.

I should also add my nodes are not all in one location or on one 
switch.  This caused issues early on with corosync.  We've consistently 
run zookeeper for the nodes, and have been much more stable for it.  I'm 
running three zookeeper nodes, one on each of the rackmount servers, and 
one more on one of the non-virt workstations.

On 03/13/2014 03:33 PM, Miles Fidelman wrote:
> Andrew, Thanks for the info!
>
> One follow-up if I  might:  When you say "I'm running qemu 1.7 on all 
> nodes" - does that include the storage-only machines, or just the 
> machines hosting VMs?
>
> Miles
>
> Andrew J. Hobbs wrote:
>> Should have read farther.  :)
>>
>> Our production system consists of a mish-mash of machines.  Four
>> workstations that were retired from normal use, and two actual servers.
>> I have two additional rackmount units that will be added soon. All VMs
>> run on the rackmount servers, while the workstations provide additional
>> drive space for sheepdog. Technically, two of those could host VMs but
>> as they don't have that much memory, I leave them as storage nodes.  We
>> have roughly 16TB of storage all together running on 0.7.7 (although I
>> will be upgrading to 0.8.0 today if cluster snapshot works for me), with
>> 11 server VMs virtualized on the cluster.  1.7TB actually in use. VMs
>> range from an internal only NFS home directory server for our Linux
>> labs, to two external webservers: one running drupal for our website,
>> the other running moodle for class interaction with students. VPN
>> services, puppet configuration server, etc also run on the cluster.
>>
>> I have one node that is iffy and will die every few weeks.  It's not a
>> sheepdog issue, but rather a Dell PERC i710 issue on that server.
>> Average time to rebalance 1.7TB when it fails?  Well, happened today at
>> 11am, finished by 12:45.  While I'm using this as an opportunity to do
>> an upgrade, I could have restarted sheepdog and the images should have
>> simply resumed.
>>
>> I'm running qemu 1.7 on all nodes, and we use kvm as the primary
>> technology rather than xen.  Nodes are running 13.10 Ubuntu, with
>> hugepages enabled.  I leave 4G for the host, the rest is allocated to
>> hugepages.  I originally had nodes running btrfs, until I had a hard
>> failure of btrfs on the iffy node.  I've since moved all nodes to ext4.
>>
>> All nodes are interconnected via gigabit ethernet.  The four
>> workstations have a single interface, which I bind to Open-vswitch, and
>> create tagged interfaces on.  Sheepdog runs on its own vlan with no
>> external access.  VMs then bridge onto ports they need access to,
>> whether it's internal only, external only, or a guest only network.  The
>> servers have four discrete nics, each of which is on a separate vlan.
>>
>> I've been very happy with the solution, as have the faculty here. Plans
>> are to extend the cluster in the future.
>>
>> On 03/13/2014 01:16 PM, Miles Fidelman wrote:
>>> Andrew: What's your environment look like (what's Sheepdog running on,
>>> what kinds of virtualization environment, what kinds of VMs?).
>>>
>>> Thanks!
>>>
>>> Miles
>>>
>>>
>>> Andrew J. Hobbs wrote:
>>>> While I haven't personally had to use them in this case (yet).  I'd
>>>> try the iSCSI or NBD options for sheepdog.   I routinely run
>>>> benchmarks when trying new combinations, if you do for both of these
>>>> situations, I'd be very interested in your results.
>>>>
>>>> I keep a linux image with a fully allocated disk to run benchmarks on
>>>> with the below command.
>>>>
>>>>
>>>> dd if=/dev/zero of=/test.out bs=4k count=1000000 oflag=direct
>>>>
>>>>
>>>>
>>>> On 03/13/2014 10:58 AM, Miles Fidelman wrote:
>>>>
>>>> Since I'm running Xen paravirtualized (for speed in all cases, and
>>>> for 2 servers, because they don't have
>>>> hardware virtualization available), QEMU based drivers aren't
>>>> available - hence my reasons for not
>>>> being able to use Sheepdog in the past.
>>>>
>>>> iSCSI solves the problem.  So would NFS.
>>>>
>>>> Thanks!
>>>>
>>>> Miles
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>
>>
>>
>
>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: ajhobbs.vcf
Type: text/x-vcard
Size: 353 bytes
Desc: ajhobbs.vcf
URL: <http://lists.wpkg.org/pipermail/sheepdog-users/attachments/20140313/dd00bbf6/attachment-0005.vcf>


More information about the sheepdog-users mailing list