Scaling Open vSwitch with a Computational Cache
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
Open vSwitch packet classification uses neural inference via NuevoMatch-APP, a reworked NuevoMatch that speeds training by ~1,000× through bucketizing ranges, optimized sampling and a lean C implementation. Two integrations are implemented on OVS-DPDK: an extra cache fronting the megaflow cache (≈1.9× geometric mean speedup for 500K ClassBench rules) and a full datapath replacement (≈12.3× geometric mean for 500K, up to 60K OpenFlow updates/s), evaluated with real traces and 10Gb links.
AI-written summary · about
chapters (8)
- 0:00Introduction to OVS
- 1:23Neural network scaling
- 2:36NuevoMatch algorithm
- 3:36Addressing training bottlenecks
- 5:55NuevoMatch-up improvements
- 9:15Integration design options
- 10:27Experimental evaluation
- 13:18Takeaways and future work
related talks
23mFast Software Cache Design for Network Appliances
20mCan OVS-DPDK be further optimised?
18mOpen vSwitch with AF_XDP as a userspace datapath
19mThe Easy Button to Kubernetes Networking with Open vSwitch
14mBrain-on-Switch: Towards Advanced Intelligent Network Data Plane via NN-Driven Traffic...
25mDPDK performance optimization for edge computing scenario
35mWhy OCP networking performances on VMs are suboptimal and how to improve them by 5x
18mSmartNIC DPU Enablement in OpenStack with OVN