ExoPlane: An Operating System for On-Rack Switch Resource Augmentation

on YouTube ↗19m

" summary

ExoPlane is an OS that augments a programmable switch with on-rack devices (smartNICs) and claims to let multiple stateful P4 apps share resources with low latency and high throughput. It implements a packet pinning model plus two-phase state updates and periodic sync for bounded inconsistency, compiles merged P4 programs, and uses a planner that solves an ILP with Gurobi. On a testbed with one switch and four smartNICs ExoPlane runs four apps when flows exceed ~100k, and outperforms app-pinning (app-pinning shows ~69–90% lower throughput) while avoiding transient drops when multiple externals are available.

AI-written summary · about

" chapters (23)

1. 0:00Intro
2. 0:35Two trends in in-network computing
3. 1:10Problem: Serving concurrent stateful apps on a switch
4. 2:51Case for on-rack switch resource augmentation
5. 4:10What should an "operating model" be?
6. 4:39Strawman model 1: App pinning Pin an app to one device and process packets on that device
7. 5:34Strawman model 2: Full disaggregation An app running on multi-devices and processing a packet on multi-devices
8. 5:59Candidate model: Packet pinning An app running on multi-devices and processing a packet on a single device
9. 6:21Our approach: Packet pinning + Union key-based flow management Union key: a union of key types of application objects
10. 7:21Exoplane design overview
11. 8:05Correctness under workload changes
12. 8:51Problem: Incorrect state eviction
13. 9:37Our solution: Two-phase state update
14. 10:11Synchronizing data plane-updatable states Multiple copies of an object entry can be updated at different places
15. 10:59Bounded inconsistency via periodic synchronization
16. 12:08Meeting requirements across apps
17. 12:36Finding optimal resource allocation using ILP
18. 13:34Putting it all together
19. 14:17Implementation and evaluation setup
20. 14:54Does packet-pinning model work well?
21. 15:46How does Exoplane work under dynamic workload?
22. 16:53Limitations and future work
23. 17:56Summary

" related talks

• 12mFlightplan: Dataplane Disaggregation and Placement for P4 ProgramsBoon Thau Loo, Isaac Pedisich, John Sonchack, Nik Sultana, Hans Giesen, Zhaoyang Han, Nishanth Shyamkumar, Shivani Burad, André DeHon · NSDI ’21
• 15mEnabling In-situ Programmability in Network Data Plane: From Architecture to LanguageYong Feng, Bin Liu, Ying Wan, Zhikang Chen, Haoyu Song, Wenquan Xu, Jiahao Li, Zijian Zhang, Tong Yun · 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22)
• 16mRuntime Programmable SwitchesJiarong Xing, Kuo-Feng Hsu, Matty Kadosh, Alan Lo, Yonatan Piasetzky, Arvind Krishnamurthy, Ang Chen · 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22)
• 14mModular Switch Programming Under Resource ConstraintsDavid Walker, Mina Tahmasbi Arashloo, Shir Landau Feibish, Mary Hogan, Jennifer Rexford · 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22)
• 19mNet-P4ct: Enhanced WAN Bandwidth Fair Sharing Using P4 Programmable SwitchesHongyu Wu, Yongping Tang, Haoran Chen, Mingwei Cui, Yihan Zou, Yihang Miao, Suhan Jiang, Damu Ding, Lirong Lai, Ming Gao, Rui Jiang, Shengyuan He, Anjian Chen, Jiaming Shi, Junjie Wan, Yandong Duan, Ruomin Fang, Qiao Kang, Guangrui Wu, Xiyun Xu · NSDI '26
• 17mHiP4-UPF: Towards High-Performance Comprehensive 5G User Plane Function on P4...Zhixin Wen, Guanhua Yan · USENIX ATC 2024
• 24mOmniXtend cache coherence protocol for datacenter CPUsZvonimir Z Bandić · ORConf 2019
• 17mMultitenant In-Network Acceleration with SwitchVMAlon Rashelbach, Mark Silberstein, Sajy Khashab · 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24)