I am not a Windows guy, I admit, but in our data center there are many Windows 2003 Server installations (32bit) and we are going to virtualize them.
Management's choice is VMWare vSphere 4 (ESX 4) but nevertheless I wanted to know how Sun xVM/Xen behaves compared to ESX. Forget all that nice Management tools offered by vmware, the only thing I was interested in is the performance feeling when actually using them.
The picture at the right (click on it to enlarge) shows the hardware list of the two "brothers": To the left you'll see the xVM DomU with paravirtualized disk and network drivers, to the right there's the vSphere ESX virtual machine with a paravirtualized SCSI disk device (and an IDE boot disk, you cannot boot in ESX from a paravirtualized disk).
Management's choice is VMWare vSphere 4 (ESX 4) but nevertheless I wanted to know how Sun xVM/Xen behaves compared to ESX. Forget all that nice Management tools offered by vmware, the only thing I was interested in is the performance feeling when actually using them.
The picture at the right (click on it to enlarge) shows the hardware list of the two "brothers": To the left you'll see the xVM DomU with paravirtualized disk and network drivers, to the right there's the vSphere ESX virtual machine with a paravirtualized SCSI disk device (and an IDE boot disk, you cannot boot in ESX from a paravirtualized disk).
Preparation
The two "machines" are prepared as follows:
Hardware and Dom0/Hypervisor:
Sun xVM:
Logical Domain/Virtual Machine:
Sun xVM HVM:
VMWare ESX:
The test program used is "Performance Test 7.0" from Passmark Software.
1. CPU
VMWare's memory driver is much faster when writing and reading large RAM blocks. I did not expect this. I ran the test several times and the results were the same.
3. Disk
Apart from the Sequential Write case it's not a surprise that xVM behaves better than the vmware PV disk. As you know writes to a ZFS storage are never "Random", the ZIL aways gets written out sequentially. So only reads can be really random - in terms of the disk heads moving.
PS: Never use the emulated IDE or SCSI devices. In xVM/Xen this is really slow and in VMWare you'll lose approx 50-66% of the performance compared to a PV disk. For this test I instructed the Performance Test program to use the PV disk partition/File system (in the vmware case because the boot volume is emulated IDE).
4. CD/DVD
This test is not really important as the ability to access mounted CDs or DVDs (as ISO images) is used rarely besides of the basic install of the operating system. But even that is done via cloning and network boot systems nowadays.
Both ISO images which were "connected" to the virtual DVD drive have been copied to the Servers: Domain-0 in the Sun XVM case, ESX datastore for the VMWare case.
The two "machines" are prepared as follows:
Hardware and Dom0/Hypervisor:
Sun xVM:
- Sun X4200M2 (2x2 core AMD Opteron 2216)
- 8 GB RAM
- 4x 135 GB SAS disk, two mirrored volumes (boot through LSI RAID mirror, xvm volume as ZFS mirror)
- Sun xVM: Xen 3.4.2-rc1-pre-xvm-
- OpenSolaris snv_121 as Domain-0, 1GB reserved for Domain-0
- Sun X4200M2 (2x2 core AMD Opteron 2216)
- 8 GB RAM
- 4x 135 GB SAS disk, two mirrored volumes (both through the LSI RAID chip)
- VMWare ESX version 4.0.0-164009
- Administered through vSphere client
- ESX consumed approx 1,1 GB itself.
Logical Domain/Virtual Machine:
Sun xVM HVM:
- A zfs block volume device is used as disk:
zfs create -b 4k -V 15G xvm/w2k3test.zvol
(NTFS uses a 4k blocksize and 15 G are sufficient for the test) - 2 GB RAM are configured.
- 2 virtual CPUs configured.
- Windows 2003 SP2 installed with all fixes/patches as Thu, Sep 10th, 2009.
- Xen Windows 2003 drivers from http://www.meadowcourt.org/downloads/
VMWare ESX:
- Datastore on a logical LSI disk device
- 15 GB virtual hard disk as IDE, 15 GB virtual hard disk as PV disk
- 2 GB RAM configured
- 2 virtual CPUs configured
- Windows 2003 SP2 installed with all fixes/patches as Thu, Sep 10th, 2009.
- VMWare tools installed.
The test program used is "Performance Test 7.0" from Passmark Software.
1. CPU
As you would expect there is not much difference between the two "candidates".
1b. 4 CPU test
To see how much overhead differs from xVM to VMWare I reconfigured the two logical machines to 4 virtual CPUs each. As there are also 4 CPUs present in hardware, the CPU performance is only affected by hypervisor CPU usage:
VMWare's overhead is bigger.
2. Memory access
1b. 4 CPU test
To see how much overhead differs from xVM to VMWare I reconfigured the two logical machines to 4 virtual CPUs each. As there are also 4 CPUs present in hardware, the CPU performance is only affected by hypervisor CPU usage:
VMWare's overhead is bigger.
2. Memory access
VMWare's memory driver is much faster when writing and reading large RAM blocks. I did not expect this. I ran the test several times and the results were the same.
3. Disk
Apart from the Sequential Write case it's not a surprise that xVM behaves better than the vmware PV disk. As you know writes to a ZFS storage are never "Random", the ZIL aways gets written out sequentially. So only reads can be really random - in terms of the disk heads moving.
PS: Never use the emulated IDE or SCSI devices. In xVM/Xen this is really slow and in VMWare you'll lose approx 50-66% of the performance compared to a PV disk. For this test I instructed the Performance Test program to use the PV disk partition/File system (in the vmware case because the boot volume is emulated IDE).
4. CD/DVD
This test is not really important as the ability to access mounted CDs or DVDs (as ISO images) is used rarely besides of the basic install of the operating system. But even that is done via cloning and network boot systems nowadays.
Both ISO images which were "connected" to the virtual DVD drive have been copied to the Servers: Domain-0 in the Sun XVM case, ESX datastore for the VMWare case.
The QEMU device emulation is slow. Really slow. It is not a good idea to work with mounted CD/DVDs in xVM's DomU.
Which version of Xen Windows GPLPV drivers you used?
0.98 from http://www.meadowcourt.org/downloads/