FBMM: Making Memory Management Extensible With Filesystems

on YouTube ↗18m

" summary

Proposes using Linux VFS callbacks to implement memory policies as FBMM file systems so memory is allocated by creating and mapping files in a mount directory and freed by deleting them. Builds standalone memory management filesystems (500–1500 LOC) for tiering, bandwidth-aware interleaving, and contiguous allocation that load separately from the kernel. Reports <8% overhead when allocating a single page and <0.1% at 128 pages, tiered MFS hits 98% of baseline throughput and sometimes outperforms kernel implementations.

AI-written summary · about

" chapters (14)

1. 0:00FBMM introduction and motivation
2. 2:29Current memory management limits
3. 3:53HugeTLBFS inspiration
4. 4:58Memory managers as filesystems
5. 7:26FBMM shim integration
6. 8:53File creation overhead optimization
7. 10:15Manual control for apps
8. 10:35FBMM shim performance
9. 11:38Basic MFS implementation
10. 12:39Tiered MFS building
11. 14:23Bandwidth MFS for NUMA
12. 15:26Contig MFS allocation
13. 16:51FBMM implementation ease
14. 17:56Conclusion and future