On Fri, Jan 11, 2013 at 8:51 AM, Luca Gervasi luca.gervasi@gmail.com wrote:

Hello,

i'm trying to create some best practices on my centos 6.3 / libvirt /kvm hypervisors.

Actually, i use NFS as shared storage backend for every VM and make reasonable use of the KSM (enabling it into qemu.conf).

If possible, use "autofs" for the NFS shares. This is because failures to connect at boot time will leave NFS unavailable for the guests, even with a simple KVM restart.

Every VM is configured with VirtIO drivers (when possible) and the disks use none as cacheing method to allow me live migration.

Review your back end storage. If it's a NetApp, especially, make sure all VM partitions are aligned for 4096 byte block boundaries. (See the NetApp white paper on this issue.) Otherwise, the NetApp is doing a lot of unnecessary re-alignment work to talk to its own back-end disks, and this can actually cause critical NetApp slowdowns.

I'll be happy to know if there are some more improvements i can implement.

Look into pair-bonding for the KVM hypervisor itself. I published notes on this at my last workplace, at https://wikis.uit.tufts.edu/confluence/display/TUSKpub/Configure+Pair+Bondin.... Also look carefully into the use of jumbo frames for NFS.

Unfortunately, that job just ended, so I'm actually looking for work in the Boston area. (If you see anything for me, let me know!)

Please share your knowledge.

Thanks in advance

Luca Gervasi