NUMA Topology & Memory Policy

NUMA architecture, memory node locality, numactl policies, and NUMA-aware tuning for latency-sensitive workloads.

Overview

NUMA (Non-Uniform Memory Access) is the memory architecture of modern multi-socket servers. Each processor socket has local RAM — access to local memory is ~60ns, access to RAM on another socket (via QPI/UPI interconnect) is ~120ns or higher.

Why NUMA Matters

A process pinned to CPU 0 that allocates memory on NUMA node 1 incurs remote memory latency on every access. At scale, this causes throughput degradation and latency spikes invisible to top/htop.

Memory Allocation Policies

local (default) — Allocate from the node where the requesting CPU runs
preferred — Prefer node N, fall back to others if full
bind — Allocate only from node N — fail if insufficient
interleave — Round-robin across nodes — maximizes bandwidth

Kubernetes and NUMA

Kubernetes Topology Manager (--topology-manager-policy=single-numa-node) ensures CPU and memory allocations for Guaranteed QoS pods are colocated on a single NUMA node, critical for latency-sensitive HPC and telco workloads. Check with:

Key concepts

NUMA node: A group of CPUs and their directly attached memory. Access to local memory is ~2x faster than remote (cross-socket via QPI/UPI) memory.
numastat: Shows per-node memory allocations and NUMA miss counters. High numa_foreign or other_node counts indicate remote memory access.
Topology Manager: Kubernetes component that aligns CPU, memory, and device (RDMA/GPU) allocations to a single NUMA node for Guaranteed QoS pods.
Interleave policy: Distributes memory pages round-robin across NUMA nodes. Trades locality for aggregate bandwidth — good for large working sets that exceed one node.

When to use it

Diagnosing unexpected latency on a multi-socket server despite low overall CPU/memory usage
Tuning database servers (Redis, PostgreSQL, MySQL) for maximum memory throughput
Explaining Kubernetes Topology Manager for latency-sensitive Guaranteed pods

Interview questions

A Redis instance on a 2-socket server shows unexpectedly high latency at moderate load. How does NUMA factor in and how do you diagnose?