NUMA — Non-Uniform Memory Access
Definition
A memory architecture where each CPU socket has its own dedicated memory controller and local memory, with access to remote (other-socket) memory available at higher latency.
Context & Usage
In modern dual-socket and multi-socket servers, each CPU has direct memory channels. Memory attached to CPU 0 is "local" for processes running on CPU 0; memory attached to CPU 1 is "remote" with ~50-80% latency overhead. Operating systems and hypervisors (VMware, Hyper-V, Linux) use NUMA-aware scheduling to pin processes to the CPU containing their memory. NUMA misconfiguration is a common cause of unexpected performance degradation in virtualization.
Examples
- Dell R750 with dual Xeon Gold 6326 has 2 NUMA nodes
- Lenovo SR650 V3 with dual EPYC 9354 has 2 NUMA nodes (or 8 with NUMA-per-CCD)
Looking for NUMA hardware?
Pro Disk Network stocks enterprise IT hardware spanning every category. Email sales@prodisknetwork.com with your requirements for a quote.