I've currently a HP DL 380 G6 server which I can use for some benchmarks I want to share with you. Maybe someone find it interesting or usefull. Both host and guest running Gentoo with kernel 2.6.31.1 (which is stable 2.6.31 update 1). The host have the following components: 2 x Intel Xeon CPU L5520 - 2 Quad-Processors (static performance, VT-d, Hyperthreading enabled in BIOS) 8 x 300 GB SAS 10k hard drives (RAID 10) 24 GB RAM (Some) Host (kernel) settings: I/O scheduler: deadline Filesystem: xfs CONFIG_HAVE_KVM=y CONFIG_HAVE_KVM_IRQCHIP=y CONFIG_KVM=m CONFIG_KVM_INTEL=m CONFIG_VIRTIO_BLK=m CONFIG_VIRTIO_NET=m CONFIG_VIRTIO_CONSOLE=m CONFIG_HW_RANDOM_VIRTIO=m CONFIG_VIRTIO=m CONFIG_VIRTIO_RING=m CONFIG_VIRTIO_PCI=m CONFIG_VIRTIO_BALLOON=m (Some) Guest (kernel) settings: I/O scheduler: deadline/cfq (see below) Filesystem: ext3 (datamode=ordered/writeback [see below)) VIRTIO network and block driver used Guest is a qcow2-image (which was expanded with "dd" to be big enough for the IO tests so that growing the image doesn't happen during the IO test. CONFIG_KVM_CLOCK=y CONFIG_KVM_GUEST=y CONFIG_HAVE_KVM=y CONFIG_HAVE_KVM_IRQCHIP=y CONFIG_VIRTIO_BLK=y CONFIG_VIRTIO_NET=m CONFIG_VIRTIO_CONSOLE=y CONFIG_HW_RANDOM_VIRTIO=y CONFIG_VIRTIO=y CONFIG_VIRTIO_RING=y CONFIG_VIRTIO_PCI=y CONFIG_VIRTIO_BALLOON=y CONFIG_PARAVIRT_SPINLOCKS=y KVM startup options (the imported ones): -m <variable-see_below> -smp <variable-see_below> -cpu host -daemonize -drive file=/data/kvm/kvmimages/gfs1.qcow2,if=virtio,boot=on -net nic,vlan=104,model=virtio,macaddr=00:ff:48:23:45:4b -net tap,vlan=104,ifname=tap.b.gfs1,script=no -net nic,vlan=96,model=virtio,macaddr=00:ff:48:23:45:4d -net tap,vlan=96,ifname=tap.f.gfs1,script=no Since we use KVM mostly for webserver stuff I've run tests with Apachebench, Graphics Magick and IOzone to simulate our workload as good as we can. Apachebench with 4 GB RAM, CFQ scheduler, 1/2/4/8 vProcs compared with the host (the hp380g6 graph): http://www.tauceti.net/kvm-benchmarks/merge-3428/ Apachebench with 2 GB RAM, CFQ scheduler, 1/2/4/8 vProcs compared with the host (the hp380g6 graph): http://www.tauceti.net/kvm-benchmarks/merge-6053/ I was interested if more RAM will give more throughput and if more vProcs will scale. As you can see 8 vProcs (compared to 2xQuadCore+HT) will give almost the same requests per second in KVM as the host alone can deliver. RAM doesn't matter in this case. Graphics Magick resize with 4 GB RAM, CFQ scheduler, 1/2/4/8 vProcs compared with the host (the hp380g6 graph): http://www.tauceti.net/kvm-benchmarks/merge-5214/ Graphics Magick resize with 2 GB RAM, CFQ scheduler, 1/2/4/8 vProcs compared with the host (the hp380g6 graph): http://www.tauceti.net/kvm-benchmarks/merge-7186/ With 8 vProcs the KVM runs about 10% slower. RAM doesn't seem to matter again here. The following IO tests run with the KVM option "cache=none": IOzone write test with 2 GB RAM, CFQ scheduler, datamode=ordered, 1/2/4/8 vProcs compared with the host (the hp380g6 graph): http://www.tauceti.net/kvm-benchmarks/merge-3564/ IOzone write test with 2 GB RAM, deadline scheduler, datamode=ordered, 1/2/4/8 vProcs compared with the host (the hp380g6 graph): http://www.tauceti.net/kvm-benchmarks/merge-4533/ It's interesting to see that the deadline scheduler gives you about 10-15 MByte/s more throughput. Please note that I haven't checked the CPU usage during the IO tests. The following IO tests run with no "cache" option set so this defaults to "writethrough": IOzone write test with 2 GB RAM, deadline scheduler, datamode ordered and writeback (see testpage), 8 vProcs compared with the hoste (the hp380g6 graph): http://www.tauceti.net/kvm-benchmarks/merge-7526/ All in all for this tests it seems that a KVM can perform almost as good as the host. I haven't tested the network throughput but it works quite well if I copy big files with ssh and/or rsync but I have no numbers yet. I'm using KVM for 2 years now and 1 year in production and I'm satisfied with it. Thanks to all developers for their work! But what is still confusing to me is the IO thing. I've had two times where new releases due to a bug destroyed some data but it was not obvious to me that it was a bug and if it was in KVM (with new kernel release) or Qemu. So I've decided to start KVM's always with "cache=none" and use ext3 as filesystem with datamode=ordered. Now the benchmark above shows quite impressive results with the KVM default option "cache=writethrough" but due to the things happend to me in the past I really fear very much to switch from "cache=none" to "cache=writethrough" because of data integrity. I know that there was some discussion some time ago but can some of the developers confirm that the following "combination" saves the data integrity of the guest filesystems and the integrity of the qcow2 image as a whole in case of a crash e.g.: Host: Filessystem xfs for KVM qcow2 images AND Guest: cache=writethrough, virtio block driver, Filesystems ext3 with datamode=ordered OR Guest: cache=none, virtio block driver, Filesystems ext3 with datamode=ordered I still don't trust ext3/ext4 with datamode=writeback or xfs as guest filesystems because of more aggressive write caching and the worst thing that can happen is that you loose data... Thanks - Robert -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
