Nested x86 virtualization under kvm is quite slow. In part, this is because nested virtualization is just slow by its very nature, and there is very little in the way of hardware acceleration to help improve its performance. However, part of the problem is that development has been (rightly) focused on correctness rather than performance. Now that nested x86 virtualization in kvm is a bit more mature, it is time to start looking for the low-hanging fruit to improve performance.
We will begin by comparing timings of primitive VMX operations under kvm and under VMware Player. Next, we will look at some macro benchmarks, and we will compare the number of VM-exits with kvm under kvm to the number of VM-exits with kvm under VMware player. Finally, we will look at runtime profiling information to examine the nested virtualization overheads of running kvm under kvm.
With apologies to AMD, this talk covers only the Intel x86 virtualization extensions.
