Non-Uniform Memory Access is what it stands for. For x86 architectures this is the method used to handle memory architectures where each processor has local memory and accessing another processor's memory is appreciably more expensive.
Non-Uniform Memory Access describes a memory architecture in which RAM is partitioned into more than one locality. Localities are called Nodes, and in most commodity hardware correlates to CPU processor sockets. In such systems access times to RAM is dependent upon which CPU is calling the FETCH and which NUMA Node the requested RAM resides in. RAM that is local to the CPU node will be fetched faster than RAM local to another CPU node.
NUMA-enabled systems provide hints to the OS in the form of certain BIOS structures. One such structure is the System Locality Information Table, which describes the relative cost of certain nodes communicating. In a fully-connected system where each node can talk directly to every other node this table is likely to have the same values for each node. In a system where nodes do not have direct connection, such as a ring topology, this table tells the OS how much longer it takes for the distant nodes to communicate.
NUMA allows NUMA-aware operating systems and programs an additional optimization center. Such programs (vmware-esx is one such) will keep process-local memory on the same NUMA-node, which in turn allows for faster memory response times. For NUMA-aware operating systems operating policy is usually set for processes to be served out of a specific NUMA node's memory for as long as possible, which also restricts execution to the cores associated with that node.
For systems that will not be running NUMA-aware programs the differential memory access times can cause seemingly undiagnosable performance differences. The severity of this disparity is very dependent upon the Operating System being used. Because of this, most server manufacturers have a BIOS option to interleave memory between NUMA nodes to create uniform access times.
Historically, older servers (before 2011) set this BIOS setting to interleave
by default. However, advances in OS support of NUMA and CPU manufacturers inter-node connection architecture advances have change this, and such settings are increasingly set to let the OS handle memory interleaving.
For Linux operating systems the command numactl
can be used to manage the memory policy for a NUMA-enabled system.