With only 4 physical vmhost servers, you should be able to tune
achieve good performance with NFS, even if you run an NFS server in a
regular linux node. For best performance, the NFS server should be on
a bare-metal OS (not inside a VM), so that it can shovel bytes from
disk to network as quickly as possible. You can experiment with
some of the fancier filesystems, but performance essentially depends
on how well the NFS server can handle the physical spindles you have.
You can handle physical storage with regular LVM tools, and watch
your network bandwidth between the nodes. Use copy-on-write VM
clones (qcow2), so the NFS server can cache disk blocks from the
master image to serve them out to many VMs.
