On 11/01/2011 01:06 AM, Matthew Law wrote: > On Mon, October 31, 2011 4:18 pm, Miles Fidelman wrote: >> Thanks guys. (FYI - I currently run DRBD over a pair of nodes that each > run md raid 10 - as I go to 4 nodes, I've been looking at different > ways of establishing a reliable storage environment that allows VMs to > migrate.) > > I think I have a similar requirement. We currently run customer VMs on > Xen 3.4 host running a mixture of MD RAID and hardware RAID cards. We > haven't had a great experience with the I/O performance of standalone KVM > VMs compared to Xen on the same hardware and a debian stable base, so at > the moment we would like to aim for Xen with the VM images on 'some kind > of clustered storage'. > Have you ever tried upstream KVM? AFAI, virtio-io in the KVM achieves pratically the same performance as XEN, even better scalability in a single host with multiple guests, from our some old tests about IO on RHEL 6. Anyway, it is not hard to implement a XEN device that redirect IO to sheepdog cluster as previously discussed, but we have much more urgen stuff to push into sheepdog and make it stable ready for production use. Thanks, Yuan > I have high hopes that we might be able to achieve this one day with xen, > ganeti and sheepdog. I am unsure of the pro's and con's of sheepdog over, > say, Ceph or Gluster for such a use case but I do like the way that > sheepdog works and it's relative simplicity (I really don't want to go > anywhere near iSCSI or ATAoE ever again!). If kvm's I/O issues can be > isolated and fixed for our use case then for me it would come down to > integrating sheepdog with ganeti. > > I am already looking at integrating sheepdog and ganeti as well as writing > some test cases for sheepdog/collie but the workload on the run-up to > Christmas is getting in the way a little. > > > Cheers, > > Matt. > > > |