Database Reference
In-Depth Information
Throughout this chapter, we have talked about how both the hardware vendors and the
software vendors have found ways to introduce memory to speed up processing. NUMA
is another example of that trend of introducing memory into the stack to speed
performance.
Today's CPUs are faster than the main memory they use. Therefore, these CPUs can
become stalled due to a lack of memory bandwidth while waiting for data they needed
to arrive from main memory. To prevent this from happening, the vendors started
creating a separate array of memory associated with each CPU socket to prevent the
performance hit associated with several processors attempting to access the same
memory at the same time.
In
Figure 7.11
, we illustrate a four-CPU system with six sockets each, and a cache of
memory associated with each physical CPU socket. Each one of those memory caches
linked to a physical CPU socket in this illustration is a NUMA node.
Figure 7.11
NUMA nodes: a four-CPU system with six cores.
In a perfect world, you could size each virtual machine to fit within a single NUMA
node and you would have optimal performance. When a CPU needs to access memory
