What does a hypervisor do?
A hypervisor is one of two main ways to virtualize a computing environment. By ‘virtualize’, we mean to divide the resources (CPU, RAM etc.) of the physical computing environment (known as a host) into several smaller independent ‘virtual machines’ known as guests. Each guest can run its own operating system, to which it appears the virtual machine has its own CPU and RAM, i.e. it appears as if it has its own physical machine even though it does not. To do this efficiently, it requires support from the underlying processor (a feature called VT-x on Intel, and AMD-V on AMD). One of the key functions a hypervisor provides is isolation, meaning that a guest cannot affect the operation of the host or any other guest, even if it crashes. As such, the hypervisor must carefully emulate the hardware of a physical machine, and (except under carefully controlled circumstances), prevent access by a guest to the real hardware. How the hypervisor does this is a key determinant of virtual machine performance. But because emulating real hardware can be slow, hypervisors often provide special drivers, so called ‘paravirtualized drivers’ or ‘PV drivers’, such that virtual disks and network cards can be represented to the guest as if they were a new piece of hardware, using an interface optimized for the hypervisor. These PV drivers are operating system and (often) hypervisor specific. Use of PV drivers can speed up performance by an order of magnitude, and are also a key determinant to performance.
Type 1 and Type 2 hypervisors – appearances can be deceptive
TYPE 1 native (bare metal)
A Type 1 hypervisor (sometimes called a ‘Bare Metal’ hypervisor) runs directly on top of the physical hardware. Each guest operating system runs atop the hypervisor. Xen is perhaps the canonical example. One or more guests may be designated as special in some way (in Xen this is called ‘dom-0’) and afforded privileged control over the hypervisor.
TYPE 2 (hosted)
A Type 2 hypervisor (sometimes called a ‘Hosted’ hypervisor) runs inside an operating system which in turn runs on the physical hardware. Each guest operating system then runs atop the hypervisor. Desktop virtualization systems often work in this manner. A common perception is that Type 1 hypervisors will perform better than Type 2 hypervisors because a Type 1 hypervisor avoids the overhead of the host operating system when accessing physical resources. This is too simplistic an analysis. For instance, at first glance, KVM is launched as a process on a host Linux operating system, so appears to be a Type 2 hypervisor. In fact, the process launched merely gives access to a limited number of resources through the host operating system, and most performance sensitive tasks are performed by a kernel module which has direct access to the hardware. Hyper-V is often thought of as a Type 2 hypervisor because of its management through the Windows GUI; however, in reality, a hypervisor layer is loaded beneath the host operating system. Another wrinkle is that the term ‘bare metal’ (often used to signify a Type 1 hypervisor) is often used to refer to a hypervisor that loads (with or without a small embedded host operating system, and whether or not technically a Type 1 hypervisor) without installation on an existing platform, rather like an appliance. VMware describes ESXi as a ‘bare metal’ hypervisor in this context.
Hypervisors versus Containers
It is mentioned that a hypervisor was one of two main ways to segment a physical machine into multiple virtual machines; the other significant method is to use containers. A hypervisor segments the hardware by allowing multiple guest operating systems to run on top of it. In a container system, the host operating is itself divided into multiple containers, each running a virtual machine. Each virtual machine thus not only shares a single type of operating system, but also a single instance of an operating system (or at least a single instance of a kernel).
Virtualization using containers
Virtualization using hypervisors
Containers have the advantage of providing lower overhead (and thus increased virtual machine density), and are often more efficient particularly in high I/O environments. However, they restrict guest operating systems to those run by the host (it is not possible, for instance, to run Windows inside a container on a Linux operating system), and the isolation between virtual machines is in general poorer. Further, if a guest manages to crash its operating system (for instance due to a bug in the Linux kernel), this can affect the entire host, because the operating system is shared between all guests.
Considerations when selecting a hypervisor
Clearly from the above, the performance and maturity of the hypervisor are going to be important considerations.
Four hypervisors under review
KVM is a Linux based open source hypervisor. First introduced into the Linux kernel in February 2007, it is now a mature hypervisor and is probably the most widely deployed open source hypervisor in an open source environment. KVM is used in products such as Redhat Enterprise Virtualization (RHEV).
Xen is an open source hypervisor which originated in a 2003 Cambridge University research project. It runs on Linux (though being a Type 1 hypervisor, more properly one might say that its dom0 host runs on Linux, which in turn runs on Xen). It was originally supported by XenSource Inc, which was acquired by Citrix Inc in 2007.
VMware is (as previously trailed) not a hypervisor, but the name of a company, VMware Inc. VMware’s hypervisor is very mature and extremely stable. It is a trusted brand that delivers excellent performance in terms of running servers, though on most loads the difference between VMware and other hypervisors is not huge. The performance (for instance time to create or start a server), however, is in general worse than either KVM or Xen.
Hyper-V is a commercial hypervisor provided by Microsoft. Whilst excellent for running Windows, being a hypervisor it will run any operating system supported by the hardware platform.
As a commercial hypervisor, the licensee must bear the cost of licensing Hyper-V itself. However, many licensees with Windows SPLA licenses see this as included within the organizations Windows licensing costs. Further, Microsoft offers preferential pricing on guest operating systems running inside Hyper-V, which in some cases may offset this cost.
Summary
Best hypervisor for you will depend on your circumstances. Typically we find traditional hosting providers and those cloud service providers that are particularly sensitive to cost or density prefer the open source hypervisors (KVM or Xen), with KVM being the most popular. Managed service providers whose customers are sensitive to branding considerations or require the enterprise style storage integration of VMware prefer that as a hypervisor. Recently we have seen Windows focused providers use Hyper-V, normally on a subset of clusters within a multi-cluster deployment.
