Medium
4 KB vs. 2 MB pages over a 4 GB random working set
A process randomly accesses a 4 GB working set. With standard 4 KB pages, that footprint requires roughly 1,000,000 distinct page-table mappings, while a typical L1 DTLB holds only around 64 entries (and a unified L2 TLB perhaps ~1,500). The workload suffers heavy, constant TLB misses, each triggering a multi-level page-table walk. Switching the same 4 GB region to 2 MB huge pages reduces it to roughly 2,048 mappings. Why does this reduce TLB misses so dramatically?
19 of 47 in Computer Architecture