WIC: Hiding Producer-Consumer Synchronization Delays with Warp-Level...

on YouTube ↗20m

" summary

WIC replaces consumer-side polling with warp-level interrupts that preempt spinning warps and park them in the UVM host kernel, using UVM page fault monitoring to detect producer data arrival and an activator to resume warps just-in-time. Implemented inside UVM with no program changes, it hides synchronization behind useful work, reduces consumer overhead across 10 evaluated applications, scales to large GPU workloads, and improves end-to-end communication by about 1.13× on average.

AI-written summary · about

" chapters (15)

1. 0:00WIC introduction
2. 1:04GPU communication challenges
3. 2:32Current sync limits
4. 4:48GPU comm analysis
5. 7:00Consumer overhead bottleneck
6. 8:24Polling fails on GPUs
7. 9:52Warp interrupt concept
8. 12:03UVM implementation
9. 13:08WIC workflow
10. 15:17Performance evaluation
11. 16:19WIC performance gains
12. 17:03Hiding sync delays
13. 17:45Scalability analysis
14. 18:48Comparison with systems
15. 19:10Conclusion