NUMA on Linux: AMD vs Intel Memory Locality (numactl)

NUMA AMD Intel linux - custom-numactl-featured.png

NUMA (non-uniform memory access) is where AMD vs Intel comparison stops being marketing and becomes milliseconds. An EPYC with multiple NUMA nodes per socket punishes a database that allocates on node 2 and runs on node 0. Intel dual-socket boxes have classic socket-wide NUMA — same toolset, different node map.

See: lscpu topology,
Proxmox NUMA pinning,
vmstat.

Print topology

numactl --hardware lists nodes, free memory, and the distance matrix.

bashhardware
numactl --hardware
lscpu | grep -i numa
numactl --hardware
nodes

Bind CPU and memory

Run latency-sensitive processes with both CPU and RAM on the same node.

bashbind
numactl --cpunodebind=0 --membind=0 postgres -D /var/lib/pgsql/data
numactl --show
numactl
bind

AMD EPYC pattern

Many hosts show multiple nodes per socket — align VM vCPU count with a node or use host passthrough planning in KVM/Proxmox.

NoteDo not enable “interleave all” unless you understand the workload — it spreads memory and hides locality problems until latency spikes.
EPYC
CCDs

Intel dual-socket pattern

Typically one NUMA node per physical CPU — numactl --physcpubind on a range of cores on socket 0 keeps traffic local.

bashinterleave check
numastat
numastat -p $(pgrep -n mysqld)
Xeon
socket

Quick Reference

  • Map: numactl --hardware
  • Run local: --cpunodebind + --membind
  • Monitor: numastat -p PID

Related tutorials

Diagrams are original illustrations by Gnome IT Solutions. Tutorial text © Gnome IT Solutions.