Axon: Fine-Grained Structural Evolution of Decision Trees in the Data Plane for Adaptive In-Network Intelligence
In-network decision-tree (DT) inference has become a practical way to execute packet-path decisions for traffic classification, security analytics, and resource steering. However, existing systems largely treat the deployed tree structure as immutable. While they can readily update parameter thresholds, structural mutations such as a leaf split or subtree prune necessitate hardware recompilation or pre-reserved layout envelopes, thus creating a stale window for adaptive workloads. To target this missing update path, we propose Axon, a runtime substrate and management system that enables fine-grained, packet-consistent structural DT evolution in the data plane without recompiling the switch program. By treating tree topology as runtime data, Axon couples a fixed RMT-compatible execution shell with a DT-aware control plane that calculates the minimal structural diff required for a new tree version. It maps this diff under strict pipeline capacity constraints and guarantees a hitless deployment. Evaluations demonstrate that Axon completely bypasses the minutes-long compilation wall, reducing structural reconfiguration latency to millisecond scale. For representative continuous updates, Axon's incremental activation yields a 5.9× speedup over the full-rule-rewrite baseline, while its proactive slack management and instruction fusion sustain line-rate throughput without memory bloating or packet drops.
Authors
- Keqin Li (ORCID: https://orcid.org/0000-0001-5224-4048)
- Yawen Wang (ORCID: https://orcid.org/0000-0002-3509-3273)
- Hai Jiang (ORCID: https://orcid.org/0009-0006-2855-7259)
- Jingran Wei (ORCID: https://orcid.org/0000-0002-2716-5935)
- Zhiqian Xu (ORCID: https://orcid.org/0009-0001-1446-838X)
Institutions
- State University of New York (US)
- Beijing University of Posts and Telecommunications (CN)
Publication Details
- Journal
- Proceedings of the ACM on Networking
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1145/3856958
- Primary Topic
- Network Packet Processing and Optimization
- Type
- article
- Field-Weighted Citation Impact
- 0.00