Service Function Chaining (SFC) ensures traffic follows a predefined sequence of service functions, enabling dynamic and efficient network services. Inspired by this, we propose an SFC-based architecture for Multi-hop Split Inference (MSI), where split sub-models are interpreted as service functions and their composition forms a service chain representing the global model. This architecture dynamically establishes communication paths for sub-models, ensuring efficient and adaptive execution. We design Neural Service Functions (NSFs) to execute sub-models as transparent TCP proxies and integrate them with Software Defined Networking (SDN), Segment Routing over IPv6 (SRv6), and the extended Berkeley Packet Filter (eBPF)-based SFC proxy. This integration enables efficient inference over dynamic routing while maintaining compatibility with existing applications. Through evaluations using well-known image recognition models and datasets, we show that the proposed architecture is feasible for MSI, particularly for latency-sensitive inference with small batch sizes, and supports dynamic path reconfiguration to adapt to changing network conditions.

• Publication:
  • Publisher's official version (may require a subscription)

Takanori Hara, Masahiro Sasabe, Service Function Chaining Architecture for Multi-hop Split Inference, IEEE Transactions on Machine Learning in Communications and Networking, 4, pp.1389-1405, August 2026.