Introduction

This site aims to be a guide to the various virtualization solutions, focussing on the x86 architecture. It provides detailed technical insight into the different approaches and compares them. Focus lies on open source applications or free-of-charge software.

What is virtualization?

The short story: Virtualization allows to share the resources of a single machine (called: Host) among several separate operating systems (OS). Resources in this context include the CPU, memory, system level hardware (like timers and interrupt controllers) and I/O devices like network cards or storage devices (hard disks). The virtualiation core software (called: Hypervisor) provides the running OSes an abstraction which allows them to run concurrently at the same time on one machine.

What is not virtualization?

The term virtualization may be used in conjuction with other terms, this is how they differ:

Emulation:The target machine instructions are interpreted one by one, allowing running software for one machine (or architecture) on a different machine (like PowerPC code on x86, x86 code on a Sparc). This is usually quite costly and thus slow. Some performance issue can be overcome by using Just-in-Time compilation afterwards, but generally you should ook at real virtualization if you don't need to run on another architecture. Most prominent application here is QEMU. Advantages lie in the easy setup, since most emulators are pure userspace applications. Security is also better, since every command is under the control of the software and a crash will only affect the emulator, not the whole host machine.
Simulation: is used to rebuild a system which is as close as possible to the original hardware. Simulators can be even cycle-accurate, providing detailed timing information. Virtualization does not care about the instruction timing, it aims to run as fast as possible. Some instructions may need much longer than others. One representative of this is SimNow.
Container virtualization: Often called operating system level virtualization. All guests run under one OS kernel (which is sometimes a limitation). The view of resources of other guests is limited (like processes or open files). Usually the file system is chroot'ed. Very fast execution since virtually no overhead. Isolation between the guests is usually limited. Prominent system here is OpenVZ.

History of Virtualization

Although Virtualization seems to be a more recent IT topic, the foundations go back into the 1970s with IBM hardware. To have a reason to sell large computers to the customers, they provides means to partition their large mainframes into several smaller units which ran independent instances of the operating system. Virtualization support was in the whole hardware, not only in the CPU.

Talks (currently in German only)

Chemnitzer Linux-Tage

Themenabend Virtualisierung (Chaos Computer Club Dresden (C3D2))

Chaos Radio Express Podcast (CRE092): Virtualisierung

more infos and download

Benchmarks

The following benchmarks provide an overview of the performance you can expect from the various solutions in various scenarios. Since serious benchmarking is a very complicated topic, please keep the following items in mind:

Most of the benchmarks are relative to native performance. Running the benchmark without any virtualization with an operation system running natively and exclusively on the machine sets the base line to 100%. The same machine is then running several (ideally all) virtualization tools while keeping most of the parameters the same (same software stack, identical operations system version, ...). The performance is the compared to the native one and the percentage is shown.
Virtualization performance depends very much on your usage scenario. The numbers are only valid for this specific benchmark. Other uses, even when similar to the one in the benchmark, can show different (even dramatically) results. I will try to describe what subsystems the benchmarks stress so you should get an impression which benchmark comes close to your application.
Although virtualization tries to abstract from the host hardware, the benchmarks can be sensitive to used machine. Relative benchmarking should sort out most of the problems, but architectural differences (like memory bandwith, cache performance) and different hardware virtualization extensions can results in different numbers on your machine:

So be warned: Your mileage may vary...

pure CPU usage: Xen and KVM above 98%, qemu around 40%, kqemu around 95%

Stressing the FPU by calculating Pi by using a taylor sum approximation (Linux source code). Please don't use this if you want to compute the value of Pi, this one converges very slowly, but does a lot of self contained computations using only very little memory. This should show a performance very close to 100% on any real virtualization solution (since only user space code without impact on memory performance).

povray numbers: Xen and KVM above 95%, qemu & kqemu not measured

Stressing the FPU and the memory interface by rendering the benchmark scene (benchmark.inc) provided with the Persistence of Vision raytracer. There should be no I/O (the resulting image is discarded), but a high memory usage.

kernbench numbers: Xen around 80%, KVM 60%, pure qemu < 5%, kqemu not measured

kernbench stresses the system by compiling a Linux kernel. This loads the CPU, has a lot of memory accesses and does I/O to load source files and headers and to store the resulting object files. Besides that many processes are created and destroyed, resulting in very frequent system calls. The numbers were gathered while compiling a stock 2.6.25.3 Linux kernel with

make allnoconfig

on a Linux RAM disk (tmpfs) (to avoid real hardware I/O). The kernel was compiled five times in a row (plus a warmup run) and was only using a single job:

kernbench -n 5 -s -H -O -M