Revert to this revision

This table compares the features and performance of the various virtualization technologies available for Linux. Hopefully this table also explains why many Linux distributions today ship Xen, even though UML and KVM are upstream and lhype is getting pretty close to inclusion into the upstream kernel.

For an explanation of the technologies, please see the [:TechOverview:technology overview] page.

If you spot something that is not up to date, or think of something missing, feel free to update this page.

full virt paravirt containers (OS virt) license architectures performance SMP guests CPU / memory hotplug standalone host notes
Xen (./) (./) GPL i686, x86-64, IA64, PPC paravirt very fast, full virt medium (./) (./) (./) full virt needs VT / AMD-V
KVM (./) (./) GPL i686, x86-64 paravirt very fast, full virt medium full and para virt need VT / AMD-V
lguest (./) GPL i686 slow/medium
rhype (./) GPL i686, x86-64, PPC fast (?) (./) research project
MoL (./) GPL PPC fast 32 bit only
UML (./) GPL i686, x86-64, PPC slow upstream
L4Linux (./) GPL i686, ARM medium (./)
qemu (./) GPL i686, x86-64, IA64, PPC slow/medium, fast with kQEMU
OpenVZ (./) GPL i686, x86-64, IA64, PPC, SPARC native (./) (./) live migration
Linux-VServer (./) GPL i686, x86-64, IA64, PPC native (./) (./) poor performance isolation
VMware (./) proprietary i686, x86-64 medium (./)
LPAR (./) proprietary s390 native (./) (./)
z/VM (./) (./) proprietary s390 very fast (./) (./) typically runs under LPAR
PHYP (./) proprietary PPC fast (./) (./) (./) used on all modern IBM System p
lv1 (./) proprietary PPC fast (./) (./) used on Sony PS3
BEAT (./) proprietary PPC fast (?) (./) used on Toshiba CellEB

Notes:

  1. Paravirtualization is fundamentally faster than full virtualization, with the exception of the userspace implementation in UML.
  2. Containers (OS-level virtualization) is yet faster than paravirtualization, achieving the native speed.
  3. Performance can vary wildly depending on workload. This page assumes system call intensive applications, since "fair weather" performance numbers are not very useful.
  4. Memory and CPU hotplug is mostly useful because it allows one to run more virtual machines on a system simultaneously, adjusting the amount of memory allocated to each guest depending on load.
  5. For an overview of the other benefits of paravirtualization, see ParavirtBenefits.
  6. Full virtualization performance in KVM and Xen is largely limited by the overhead of trap & emulate. Emulating multiple instructions at once at the time of a trap should bring it up to speed with VMware.
  7. OpenVZ (Virtuozo) and VServer are not virtualization technologies per se. They carve up a single system in "super chroot" jails. All the containers run on top of the same kernel.
  8. OpenVZ (Virtuozo) can change memory and CPU quota during runtime, there is no real hotplug since there are no guest kernels.
  9. Qemu can emulate different guest architectures, eg. running an x86 virtual machine on a PPC guest. Qemu also has the distinction of being the only full virtualization technology that can run without root privileges.
  10. Parts of Qemu are used in the full virtualization implementations of Xen and KVM.
  11. The standalone host column indicates whether or not the hypervisor (or host OS, in the case of VM) is booted before Linux. See the [:HypervisorVsLinuxBased:comparison of hypervisor based vs. Linux based virtualization] for the debate on whether or not this is an advantage.