Skip to main content
Definition Website Computer and Technology

Non-Uniform Memory Access

Subject: Computer and Technology noun /ˌnɑnˈjuː.nɪ.fɔːrm ˈmɛm.ə.ri ˈæk.sɛs/ 1 definition

  1. Definition (noun)

    Entry #201 · Updated

    Non-Uniform Memory Access (NUMA) is a shared-memory multiprocessing architecture in which each processor node has directly attached local memory, making access latency and bandwidth depend on where data resides relative to the requesting core. While all nodes share a single physical address space, reading or writing local memory is significantly faster than traversing an inter-socket interconnect to reach remote memory on another node. Unlike Uniform Memory Access (UMA), where a single shared memory bus creates contention as core counts grow, NUMA scales aggregate memory bandwidth across nodes at the cost of requiring NUMA-aware operating system placement and scheduling to avoid remote-access penalties.

    Common Forms

    • Cache-Coherent NUMA (ccNUMA): Hardware directory or snoop protocols automatically keep processor caches synchronized across nodes, allowing standard multithreaded applications to run transparently on multi-socket servers.

    • Non-Cache-Coherent NUMA (nccNUMA): Processors directly address remote memory without hardware cache-coherence tracking across nodes, shifting data synchronization to software to simplify specialized high-performance interconnects.

    • Sub-Socket NUMA: High-core-count processors partition a single physical socket into multiple logical NUMA domains aligned with on-die memory controllers or chiplets, such as Intel Sub-NUMA Clustering or AMD multi-node-per-socket modes, to lower local access latency.

    • Virtual NUMA (vNUMA): Hypervisors expose physical or synthetic NUMA topologies to guest virtual machines so guest operating systems can co-locate threads with their allocated memory pages.

    Example

    On a dual-socket server with 128 GB of RAM per socket, an in-memory analytics worker running on Node 0 accidentally allocates its working buffers on Node 1, forcing cache misses across the inter-socket interconnect and sharply degrading throughput. Launching the job with numactl --cpunodebind=0 --membind=0 pins both its threads and memory pages to Node 0, eliminating cross-socket traffic and restoring local bandwidth.