NUMA

Non-uniform memory access (NUMA) systems expose locality-dependent memory and I/O costs: a processor accesses local memory and devices differently from resources attached to another socket or node.

核心思想

NUMA topology makes placement part of performance correctness. Threads, pages, queues, accelerators, and storage devices should be mapped with knowledge of socket and interconnect locality; otherwise remote traffic, cache-coherence work, and bandwidth contention can dominate nominal compute or device capability.

为什么重要

Many high-core-count and multi-device results in this corpus depend on pinning, allocation policy, and I/O placement. A benchmark that omits topology can be irreproducible or incorrectly attribute a locality effect to an algorithm.

关键观察 / 隐含假设

• 观察：memory/swap or I/O resource partitioning interacts with core locality. ScaleSwap-FAST26 and MAIO-FAST26 evaluate system paths with such constraints.
• 观察：accelerator/device placement can expose cross-node traffic. DSA-2LM-ATC25 and Catur-MLSys26 use hardware-aware execution contexts.
• 假设：pinning once is sufficient. Dynamic scheduling, page migration, shared data, and multi-tenant placement can invalidate a static mapping.

设计空间与取舍

• Locality vs load balance：strict locality can strand capacity; balancing can add remote traffic.
• First-touch, binding, or migration：allocation policy changes both steady-state bandwidth and adaptation cost.
• CPU, memory, and device topology：optimizing only one layer can move contention to another.

引用本概念的论文

• ScaleSwap-FAST26 — core-centric swap-resource management.
• MAIO-FAST26 — memory/I/O placement context.
• DSA-2LM-ATC25 — hardware-aware acceleration path.
• Catur-MLSys26 — device/runtime placement boundary.
• SoarAlto-OSDI25 — system scheduling/locality context.