The KVM hypervisor supports overcommitting CPUs and overcommitting memory. Overcommitting is allocating more virtualized CPUs or memory than there are physical resources on the system. With CPU overcommit, under-utilized virtualized servers or desktops can run on fewer servers which saves power and money.

With KVM, virtual machines are Linux processes. Guests on the KVM hypervisor do not have blocks of physical RAM assigned to them instead they function as processes. Each process is allocated memory when it requests more memory. KVM uses this to allocate memory for guests when the guest operating system requests more or less memory. The guest only uses slightly more physical memory than the virtualized operating system appears to use.

When physical memory is nearly completely used or a process is inactive for some time, Linux moves the process's memory to swap. Swap is usually a partition on a hard disk drive or solid state drive which Linux uses to extend virtual memory. Swap is significantly slower than RAM.

As KVM virtual machines are Linux processes, memory used by virtualized guests can be put into swap if the guest is idle or not in heavy use. Memory can be committed over the total size of the swap and physical RAM. This can cause issues if virtualized guests use their total RAM. Without sufficient swap space for the virtual machine processes to be swapped to the pdflush process starts. pdflush kills processes to free memory so the system does not crash. pdflush may destroy virtualized guests or other system processes which may cause file system errors and may leave virtualized guests unbootable.

The swap partition is used for swapping underused memory to the hard drive to speed up memory performance. The default size of the swap partition is calculated from amount of RAM and overcommit ratio. It is recommended to make your swap partition larger if you intend to overcommit memory with KVM. A recommended overcommit ratio is 50% (0.5). The formula used is:

(0.5 * RAM) + (overcommit ratio * RAM) = Recommended swap size

The Red Hat Knowledgebase has an article on safely and efficiently determining the size of the swap partition — refer to Knowledgebase.

It is possible to run with an overcommit ratio of ten times the number of virtualized guests over the amount of physical RAM in the system. This only works with certain application loads (for example desktop virtualization with under 100% usage). Setting overcommit ratios is not a hard formula, you must test and customize the ratio for your environment.

Virtualized CPUs are overcommitted best when each virtualized guest only has a single VCPU. The Linux scheduler is very efficient with this type of load. KVM should safely support guests with loads under 100% at a ratio of 5 VCPUs Overcommitting single VCPU virtualized guests is not an issue.

You cannot overcommit symmetric multiprocessing guests on more than the physical number of processing cores. For example a guest with four VCPUs should not be run on a host with a dual core processor. Overcommitting symmetric multiprocessing guests in over the physical number of processing cores will cause significant performance degradation.

Assigning guests VCPUs up to the number of physical cores is appropriate and works as expected. For example, running virtualized guests with four VCPUs on a quad core host. Guests with less than 100% loads should function effectively in this setup.