On Tue, Nov 10, 2009 at 11:36:39PM +0200, Pasi Kärkkäinen wrote:
On Tue, Nov 10, 2009 at 01:52:36PM -0500, Scott McClanahan wrote:
Yeah.. Xen paravirtualized mmu is fast, and in some (many) cases beats CPU hardware virtualized mmu.
KVM has 'pvmmu' aswell, but it's not as good, so KVM is faster with CPU hardware virtualization. But that's a problem of KVM only, they haven't managed to optimize the pvmmu. And they're going to drop it altogether.
KVM people tend to say 'paravirtualized mmu is slow', but they just mean KVM implementation of it sucks :)
-- Pasi
I haven't tested or seen any benchmarks but I wonder how much the addition of a page table for virtualized guests will help. Not to mention newer features like a virtualized task priority register and ASID could continue to require less paravirt code in the guest. I get my two new 5500 series servers in a few weeks so I'm pretty excited to see some of the second gen hardware virtualization assist features in action.
I don't know. Of course hardware will add features and get more optimized in the future.
Some benchmarks from IBM guys, Xen vs. KVM: http://www.spinics.net/lists/kvm/msg13910.html http://www.spinics.net/lists/kvm/msg14068.html http://www.spinics.net/lists/kvm/msg21913.html
And forgot this one: http://www.spinics.net/lists/kvm/msg16579.html
-- Pasi
Quotes:
"So, KVM requires 66.93/52.85 = 26.6% more CPU to do the same amount of work." "If we normalize to CPU utilization, Xen is doing 20% more throughput." "KVM running Windows VMs uses 46% more CPU than the Other-Hypervisor" "A different hypervisor was compared; KVM used about 60% more CPU cycles to complete the same amount of work."
I bet KVM will catch up at some point.. at the moment it seems to not perform as good as Xen. Then again it's a much younger product.