Network & Infrastructure / Linux Networking & Host Performance
NUMA locality and cross-socket networking latency
A high-throughput server shows inconsistent latency when the application moves between CPU sockets. NIC interrupts and application memory may be placed on different NUMA nodes.
Environment and inputs: Linux network namespaces, host routing and network drivers; availability and semantics depend on kernel and driver versions. Diagnostic/design example; not tested on production equipment.
Collect the evidence
Record NIC locality, interrupt affinity, application CPU affinity, memory placement and per-socket load. Compare controlled runs with consistent workload and connection count.
Explain the result
Remote memory and cross-socket processing can add overhead even when total resources appear sufficient. A global CPU graph hides locality and contention that affect networking performance.
Plan the fix
Align processing and memory placement with the supported application and NIC topology, leaving capacity for failover and other workloads. Avoid pinning every network task to one already busy socket.
Check the result and rollback
Validate throughput, tail latency and CPU headroom under realistic concurrency. Restore prior affinity if local optimization starves another workload or reduces resilience when one processing group is overloaded.