HyperSFC: State-Intensive Service Function Chaining on Hyper-Converged Edge Infrastructure

Service function chaining (SFC) enhances network efficiency and agility by enabling the orchestration of network functions (NFs) in a customized manner. In public cloud environments, SFC is implemented via sequential packet routing through a series of NFs, each hosted on an x86 server pool. However, such server pooling-based SFC provisioning is not easily deployable in the edge cloud due to constraints related to hardware budget and deployment footprints. In the edge cloud, hyper-converged infrastructure (e.g., switch servers) offers a promising solution for cost and footprint reduction. Typically, a hyper-converged switch server comprises a programmable switch and a CPU, allowing NFs to be hosted on the CPU with packet routes controlled by the switch. To enhance performance, forwarding tables of NFs can be further offloaded to the switch. However, due to the limited on-chip memories of the switch, NFs with intensive states cannot be fully offloaded. To address this, we propose busy submit via packet circulation between the switch and CPU. While busy submit ensures packets experiencing table search misses in the switch are routed to the full tables in the CPU, it may lead to repeated circulations in the context of SFC processing, increasing both packet latency and CPU load. To mitigate this, we introduce lazy submit, reducing packet circulation overhead by postponing packet submission to the CPU until the end of the pipeline. Additionally, we use 2D Bloom filters to cache NF dependencies, reducing unnecessary circulations. The evaluation of an NF chain with 10 NFs shows that busy submit enables state-intensive SFC, while lazy submit reduces overall latency by 60%.