@inproceedings{Ding:2026:TLBShootdown,
 abstract = {Translation Lookaside Buffers (TLBs) cache recent virtual-to-physical address translations. Each core has a private TLB. Because TLBs are not coherent, the OS must perform a TLB shootdown to invalidate stale translations when page-table entries change. TLB shootdowns are a major scalability bottleneck on large multicore systems. Prior work focused on avoiding expensive large-scale shootdowns by target-set reduction, deferred invalidation, batching, and early acknowledgments. However, the effect of hardware topology on TLB shootdown latency has been largely ignored.We explore the effects of the hardware topology and the broadcast structure on the TLB shootdown latency in Linux's TLB shootdown mechanism. Our results show that topology-aware propagation can reduce TLB shootdown latency for large target sets by up to 1.9{\texttimes} over Linux's default mechanism. However, no single propagation structure is universally optimal. Performance depends on the hardware topology, workload behavior, and core count. Eager removal of cores from the target set further improves performance.},
 address = {New York, NY, USA},
 author = {Ding, Ning and Seltzer, Margo and Achermann, Reto},
 booktitle = {Proceedings of the 14th Workshop on Programming Languages and Operating Systems},
 doi = {10.1145/3831586.3838155},
 id = {Ding:2026:TLBShootdown},
 isbn = {9798400728778},
 keywords = {TLB Shootdown, Multi-processor Systems, Virtual Memory},
 location = {Prague, Czech Republic},
 note = {},
 numpages = {8},
 pages = {17–24},
 publisher = {Association for Computing Machinery},
 series = {PLOS '26},
 title = {Topology-aware TLB Shootdown Protocols: It's more complicated than you think},
 url = {https://doi.org/10.1145/3831586.3838155},
 year = {2026}
}