UNUM: A New Framework for Network Control

on YouTube ↗16m

" summary

A transformer-based state embedder (UNUM) predicts bucketized future network signals rather than exact values, improving prediction accuracy by up to ~50% versus regression and enabling a shared global state used as a service. Integrations include a tiny MLP adapter or full retrain; the adapter yields practical gains: 13% Orca reward uplift and about 45% Pensieve improvement, while BBR sees similar throughput and a 22% queue delay reduction. The system runs on CPU with ~2.68 ms decision delay and preserves most control performance.

AI-written summary · about

" chapters (13)

1. 0:00UNUM framework intro
2. 1:20Current control limits
3. 3:47Better state estimation need
4. 4:50Transformer state modeling
5. 6:38Prediction accuracy classification
6. 7:40Separation principle in design
7. 9:04Integrating existing controllers
8. 10:49UNUM async implementation
9. 11:11Evaluation results
10. 12:34Context length impact
11. 13:18Parameter efficiency benefits
12. 14:21Decision delay trade-offs
13. 15:25Conclusion

" related talks

• 14mBrain-on-Switch: Towards Advanced Intelligent Network Data Plane via NN-Driven Traffic...Qi Li, Jinzhu Yan, Haotian Xu, Zhuotao Liu, Ke Xu, Mingwei Xu, Jianping Wu · 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24)
• 17mMutant: Learning Congestion Control from Existing Protocols via Online Reinforcement...Lorenzo Pappone, Alessio Sacco, Flavio Esposito · NSDI '25
• 16mCredence: Augmenting Datacenter Switch Buffer Sharing with ML PredictionsStefan Schmid, Vamsi Addanki, Maciej Pacut · 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24)
• 11mScaling Distributed Machine Learning with In-Network AggregationMarco Canini, Masoud Moshref, Changhoon Kim, Arvind Krishnamurthy, Dan R. K. Ports, Jacob Nelson, Amedeo Sapio, Chen-Yu Ho, Panos Kalnis, Peter Richtarik · NSDI ’21
• 17mDOTE: Rethinking (Predictive) WAN Traffic EngineeringChaim Hoch, Srikanth Kandula, Yarin Perry, Felipe Vieira Frujeri, Ishai Menache, Michael Schapira, Aviv Tamar · NSDI 2023
• 18mLearning-Enhanced High-Throughput Pattern Matching Based on Programmable Data PlaneMingwei Xu, Qing Li, Yong Jiang, Zili Meng, Guanglin Duan, Yucheng Huang, Zhengxin Zhang, Dan Zhao, Dirk Kutscher, Ruoyu Li · USENIX ATC '25
• 12mBreaking the Transience-Equilibrium Nexus: A New Approach to Datacenter Packet TransportMohammad Alizadeh, Anirudh Sivaraman, Shiyu Liu, Ahmad Ghalayini, Balaji Prabhakar, Mendel Rosenblum · NSDI ’21
• 12mATP: In-network Aggregation for Multi-tenant LearningMichael Swift, Aditya Akella, Wenfei Wu, ChonLam Lao, Kshiteej Mahajan, Yanfang Le, Yixi Chen · NSDI ’21