ShapeShifter: Cross-Model Fast Reconfiguration for In-network ML Classification
In-network machine learning enables low-latency inference in programmable data planes, enabling real-time traffic analysis and line-rate machine learning classification. However, existing solutions are optimized for single-user, single-model deployments, and require recompilations for cross-model updates. Faster and consistent updates between model types would greatly benefit ML developers and cloud operators who want to use this technology in production. To this end, we propose ShapeShifter, an end-to-end system for fast cross-model reconfiguration of in-network ML inference. ShapeShifter unifies the inference workflow of different ML classifiers into three primitives: compute, sign extract, and vote. ShapeShifter uses a block-based design to enable cross-model resource sharing and improve resource flexibility. To reduce reconfiguration time, ShapeShifter's control plane adopts a differential reconfiguration algorithm to reduce the number of atomic operations. Our evaluation on real hardware shows that ShapeShifter supports runtime reconfiguration across nine model types without recompiling the data-plane program, while reducing the number of update operations by 51.3% and speeding up reconfiguration by up to 51%.
Authors
- Hyojoon Kim (ORCID: https://orcid.org/0000-0002-0650-5339)
- Jianxi Chen (ORCID: https://orcid.org/0009-0008-3476-8903)
- Di Zhu (ORCID: https://orcid.org/0009-0009-0924-5366)
Institutions
- University of Virginia (US)
Publication Details
- Journal
- Proceedings of the ACM on Networking
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1145/3856975
- Primary Topic
- Software-Defined Networks and 5G
- Type
- article
- Field-Weighted Citation Impact
- 0.00