Talk by Rob Johnson (VM Ware)

on YouTube ↗01h 06m

" summary

A concrete solution called mosaic pages groups k regular 4KB pages so a single TLB row covers multiple pages while avoiding physical contiguity by compressing per-page frame pointers down to ~log log bits using a low‑associativity hash table. The new hashing primitive iceberg hashing uses a front yard (single choice) and backyard (2‑choice) to keep buckets small and stable. Simulated on gem5 (Graph500, 1GB footprint, 16B instructions) and patched Linux results show large TLB miss reductions, no I/O increase, 36→7 bit frame encodings, and practical hardware cost (tabulation hashes ~2k gates); a scalable Horizon LRU avoids global contention.

AI-written summary · about

" chapters (19)

1. 0:00Mosaic pages project intro
2. 1:26Systems research collaboration
3. 2:49Address translation overheads
4. 8:15TLB size limitations
5. 12:13Current solutions shortcomings
6. 16:12TLB improvements requirements
7. 16:33Mosaic pages concept
8. 23:27Iceberg hashing foundation
9. 27:19Power of two choices limits
10. 31:16Mosaic theorem construction
11. 36:41Theory versus systems memory
12. 40:33Co-designing models theorems
13. 40:55Memory augmentation rationale
14. 44:09Horizon LRU paging
15. 55:46Theory to practical systems
16. 57:54Hardware simulation results
17. 1:01:29Linux kernel I/O costs
18. 1:03:35TLB latency feasibility
19. 1:05:27Key takeaways research