Scalable and Effective Page-table and TLB management on NUMA Systems

on YouTube ↗17m

" summary

Hydra implements on-demand partial replication of page-table entries inside Linux on x86_64, tracking per-node sharers so page-table reads are local while churn touches only affected replicas. The prototype runs on 4- and 8-socket NUMA machines, cuts TLB shootdowns dramatically (microbenchmarks show a 15× speedup on a munmap/mprotect-style workload versus baseline), matches eager full-replication read performance, uses far less page-table memory than full replication, and yields +12% and +36% runtime on Webserver and Memcached respectively.

AI-written summary · about

" chapters (7)

1. 0:00NUMA page-table challenges
2. 2:09Page tables on NUMA systems
3. 3:58Hydro solution overview
4. 7:15TLB shootdown overhead
5. 9:51Hydro’s lazy replication
6. 13:15Performance gains
7. 15:48Conclusion