Phoronix: The Start Of Some KVM Virtualization Benchmarks
The last time we published any benchmarks using KVM (the Kernel-based Virtual Machine) virtualization was last year when looking at the performance with the Linux 2.6.31 kernel and before that when looking at the Intel Core i7 virtualization performance. However, a new set of Linux virtualization benchmarks are being worked on...
I'm not a very serious KVM user but my one complaint about it is that compiling is very slow, much more by comparison than anything else. I believe this is a known problem. It has been suggested that setting the virtio storage device to rotational mode on the guest and using the deadline scheduler on the host can help a little but Windows XP doesn't support virtio storage. I think it only gives you a small gain anyway.
It would be useful to know exactly how you configured the storage and networking of the VM, as these can obviously have a significant impact on performance. Especially, I believe you get the best disk performance if you use a host LVM volume as the storage for the VM. I can certainly put you in touch with some of the RH virt folks if it would help to make sure you get the optimal KVM configuration.
When you're doing the virtualbox testing, it'd be nice to see the results both with and without hardware virt support (seeing as how qemu's equivalent kqemu seems depreciated...).
I second (or third?) the call for virtio disk/nic testing and comparisions, and the other kernel paravirt options.
Actually, if testing with a "bleeding edge" kernel or whatever, seeing what kind of difference the new host-side vhost net driver makes would be sweet too.
Originally Posted by lordmozilla
have you got an agreement with vmware? Don't they still dissalow all (published) benchmarking in their EULA?
I second that.
KSM can have a devastating effect on the performance.
Beyond that, using all of host's memory for a VM is a very bad idea, you lose the disk caching on the host side and may even force the host to swap output memory.
Upgrade the machine to 6GB (give 4 to the guest).
Graphical test should performance better with -vga std or vmware.
Besides an upgrade to 6Gb, I would also suggest to give 50% of host memory to the guest, so that we could compare the virtual systems benchmarks (using 3Gb) with the native execution with 3Gb (an important extra).
Michael, were you considering something like this?