Non-uniform memory access, also known as NUMA, is designed with memory locality in mind so that pools of adjacent memory are placed in islands called NUMA nodes. Here, memory is divided between physical CPUs into NUMA nodes. In a dual CPU server, half the memory is on NUMA node 0 along with CPU 0, and the other half of memory is on NUMA node 1 with CPU 1. The CPU in NUMA node 0 considers the memory in NUMA node 0 to be local (fast) and the memory in NUMA node 1 to be remote (slower). Because of this, the more memory we can give a VM within a NUMA node, the more efficiently that VM can access its memory.
vSphere 6.5 has made changes to its CPU scheduler to optimize the virtual NUMA topology. In vSphere 6.0, if you were to create a VM with 16 CPUs and two cores per socket, eight Virtual Proximity Domains would be created. If you create a VM with the same configuration in vSphere 6.5, only two Virtual Proximity Domains are created, which match the Physical Proximity Domains of the host.
What if your VM needs to be bigger than a NUMA node? One of the features originally introduced in vSphere 5 is Virtual NUMA (vNUMA) or the ability of NUMA inside a VM to be presented to the guest OS.
vNUMA is designed for modern OSes that are NUMA-aware and can make intelligent page management decisions based on locality. A characteristic aspect of vNUMA is that it incorporates distributed shared memory (DSM) inside the hypervisor, which is in contrast to the more traditional approach of providing it in the middleware.
When creating a VM, you have the option to specify the number of virtual sockets and cores per virtual socket. If the number of cores per virtual socket on a vNUMA-enabled VM is set to any value other than the default of one and that value doesn't align with the underlying physical host topology, performance might be slightly reduced.
Therefore, for best performance, if a VM is to be configured with a non-default number of cores per virtual socket, that number should be an integer multiple or integer divisor of the physical NUMA node size.