OmniCache: Collaborative Caching for Near-storage Accelerators

on YouTube ↗17m

" summary

OmniCache combines a host near-cache and a 4GB device cache (example config: 16GB host + 4GB device) with an OM index and user library to keep data in the nearest cache, cut alignment write amplification, and avoid repeated 4KB fetches for 1KB reads. A model-driven offloading engine predicts transfer cost, queuing delay, and execution time to place work on host or device. Evaluation on an emulated programmable storage platform with added PCI latency shows up to 3.24× I/O and 3.06× data-processing speedups, and CXL memory mapping (≈200 ns) reduces CPU polling and queuing overhead.

AI-written summary · about

" chapters (15)

1. 0:00OmniCache introduction
2. 2:49Existing caching limits
3. 5:17OmniCache design
4. 7:02OmniCache workflow
5. 8:51Read amplification reduction
6. 9:34Horizontal caching for stalls
7. 10:18Dynamic offloading decisions
8. 11:43CXL for unified memory
9. 12:24Experimental setup
10. 13:27Macro benchmark results
11. 14:51OmniDynamic validation
12. 15:12CXL memory benefits
13. 15:32Real-world gains
14. 16:15Conclusion and future
15. 16:58Q&A