A High-Speed Stateful Packet Processing Approach for Tbps Programmable Switches

on YouTube ↗16m

" summary

RIBOSOME splits packets at a Tofino ASIC, sends only headers to a FastClick NF on one 100G NIC and stores payloads on shared servers via RDMA, achieving 300 Gbps through a single 100G NIC and a 3x throughput gain over a baseline. RDMA roundtrips add ~4 µs latency; storing ~500 headers (≈36 KB) on-switch suffices for 1 Tbps at 1 KB average. Evaluation hits 93 Mpps at 400 B packets, and runs stateful load balancers and schedulers with up to a 2.2× speedup over prior software.

AI-written summary · about

" chapters (9)

1. 0:00Network function challenges
2. 2:10Header-only processing
3. 4:23The Ribosome architecture
4. 6:01Technical implementation
5. 8:04Throughput evaluation
6. 10:28Latency performance
7. 11:55Scalability and packet size
8. 13:27Advanced function testing
9. 15:02Conclusion

" related talks

• 19mEnabling Programmable Transport Protocols in High Speed NICsDavid Walker, Manya Ghobadi, Mina Tahmasbi Arashloo, Jennifer Rexford, Alexey Lavrov, David Wentzlaff · NSDI 2020
• 47mThe World of 100G NetworkingChristoph H Lameter · LinuxCon + ContainerCon Europe 2016
• 32mSnabb Switch simple and fast packet networkingLuke Gorrie · HighLoad++ 2015
• 31mHow to capture 100G Ethernet traffic at wire speed to local diskChristoph H Lameter · linux.conf.au 2020
• 32mSnabb Switch simple and fast packet networkingLuke Gorrie · HighLoad++ 2015
• 32mSnabb Switch simple and fast packet networkingLuke Gorrie · HighLoad++ 2015
• 32mSnabb Switch simple and fast packet networkingLuke Gorrie · HighLoad++ 2015
• 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)