Wire Speed Name Lookup: A GPU-based Approach
summary
A GPU-based name-lookup engine implements a multi-ATA state table mapped to GPU memory and a pipelined CPU–GPU execution style, claiming wire-speed lookups of 63.52M searches/sec and aggregate forwarding over 100 GB/s on up to 10M-name tables, with memory compression reported as >100× versus dense tables and update rates of ~30K inserts/sec and ~600K deletes/sec; experiments show stable latency under 100 μs per packet and scalability arguments toward tens of millions of entries.
AI-written summary · about
chapters (9)
- 0:00Start
- 0:41Background & Motivation
- 2:19name lookup challenges
- 4:04Algorithm & Data Structure
- 7:34Name Lookup Engine Framework
- 8:07Latency Optimization
- 10:13i-GPU Memory Access Optimization
- 11:25Experimental Results
- 14:08Conclusion
related talks
25mKeynote
22mBlackwire: WireGuard in HDL at 100+ Gbit/s
18mArchitecting Multi-tenant Data-center Networks for GPU Customers
19mFpgaNIC: An FPGA-based Versatile 100Gb SmartNIC for GPUs
31mData Streaming from SDR to Servers for Cognitive Radar and EW
13mSNARY: A High-Performance and Generic SmartNIC-accelerated Retrieval System
27mFaster Than Fast: Networking and Communication Optimizations for Llama 3
31mLive Q+A 3
