NUMA (Non-Uniform Memory Access) is the hardware reality of multi-socket systems. Each CPU socket has its own memory controller and DIMMs; a thread on socket 0 accessing memory on socket 1 crosses the inter-socket interconnect and pays extra latency. Modern CPUs mitigate this with caches, but bandwidth-heavy HPC code suffers measurably.

Install numactl:

sudo apt install numactl           # Debian/Ubuntu
sudo dnf install numactl           # Fedora/RHEL

Discover the topology:

numactl --hardware

This shows available nodes (sockets), which CPUs belong to each node, memory per node, and the distance matrix (relative latency between nodes).

Example output:

available: 2 nodes (0-1)
node 0 cpus: 0 1 2 3 4 5 6 7
node 0 size: 64000 MB
node 1 cpus: 8 9 10 11 12 13 14 15
node 1 size: 64000 MB
node distances:
node   0   1
  0:  10  21
  1:  21  10

Distance values are relative latency costs. Local access (node 0 from cores on node 0) costs 10 units. Cross-socket (node 1 from cores on node 0) costs 21 units—roughly 2x slower. Actual latency depends on hardware; these are relative numbers.

Understanding the numbers: - available: 2 nodes (0-1) — 2 NUMA nodes (sockets), numbered 0 and 1 - node 0 cpus: 0 1 2 3 4 5 6 7 — cores 0-7 are on socket 0 - node 0 size: 64000 MB — socket 0 has 64GB of memory - Distance matrix — [i,j] entry is relative latency from node i to node j

On a 4-socket or 8-socket machine, the distances are more complex, but the principle is the same: local access is fastest, and cross-socket costs increase with distance.

Other tools:

lscpu                   # show CPU topology (simpler than NUMA view)
numastat                # show memory distribution
numastat -p PID         # show memory placement for a process
taskset                 # simpler CPU pinning (no memory control)

numactl is the comprehensive tool for NUMA control; taskset is simpler if you only care about CPU affinity. For HPC, use numactl to control both CPU and memory placement together.