In a multi-NUMA node environment, SBC performance can degrade significantly if traffic processing tasks are not properly bound to the same NUMA node. The penalties associated with inter-NUMA communication primarily arise due to increased memory access latency and cross-socket CPU communication overhead.

Sources of Inter-NUMA Node Penalties include:

• Remote Memory Access Latency

  When an SBC process running on NUMA Node 0 accesses memory allocated on NUMA Node 1, it introduces higher latency compared to local memory access.

  The latency can be 30-50% higher or more when accessing remote memory compared to local memory.

• Cross-NUMA CPU Synchronization

  If different SBC threads (e.g., signaling and media handling) are spread across NUMA nodes, the L3 cache coherence traffic increases across sockets.

  This leads to higher interconnect bandwidth usage (e.g., Intel UPI or AMD Infinity Fabric), causing potential performance bottlenecks.

• Network Interrupt Processing Across NUMA Nodes

  If a NIC is assigned to NUMA Node 0, but the SBC process handling packets is running on NUMA Node 1, then every packet transfer incurs an inter-socket latency.

  This results in degraded packet processing performance, especially in high-throughput VoIP and SIP traffic environments.