Bandwidth Speaks, We Listen: Dynamic Memory Interleaving for Tiering

on YouTube ↗28m

" summary

Design and implementation of Dynamic Memory Interleaving (DMI) actively reshuffles hot pages between local DDR and CXL remote memory using a user-space calibrator and kernel migrator to maximize aggregate bandwidth. DMI uses FFT-based interval selection to pick measurement windows, converges to stable interleave ratios in ~30–40s, and runs on a 72-core Granite Rapid testbed with 260 GB/s local and 220 GB/s CXL bandwidth. Evaluation shows DMI achieves up to 26% speedup over prior bandwidth-aware tiering and stays within ~5% of an ideal static ratio.

AI-written summary · about

" chapters (20)

1. 0:00Dynamic memory interleaving intro
2. 1:12Memory bandwidth scaling
3. 2:59Static interleaving limits
4. 4:02Dynamic approach problems
5. 5:45Design goals of DMI
6. 6:48Calibrator algorithm
7. 7:50Bandwidth measurement interval
8. 10:40FFT for adaptive measurement
9. 13:07Migrator’s page selection
10. 14:12Experimental setup
11. 14:54Performance vs baselines
12. 15:38Convergence to ideal ratio
13. 16:19Evaluating FFT interval
14. 18:06Variable bandwidth demands
15. 19:09Conclusion and future
16. 19:38Q&A: DAM caches trade-offs
17. 21:04Q&A: Bandwidth restriction
18. 22:49Q&A: Migration overhead
19. 23:54Q&A: FFT in large systems
20. 24:40Closing remarks