The piece reverse-engineers Skylake-X’s next-page prefetcher, a mechanism Intel has documented only sparsely. It explains how such a prefetcher differs from ordinary cache-line prefetching: it anticipates movement across 4 KB page boundaries and can bring future page translations into the TLB. The investigation asks whether Skylake-X has this mechanism, what access pattern activates it, how many future translations it fetches, where it places them, and whether it works in both address directions. Intel’s descriptions of Ivy Bridge and the newer Redwood Cove provide context, but leave open what happened in intervening generations.
The experiments use Linux performance counters to measure TLB misses while walking mapped pages, and first establish that Skylake-X does have a next-page prefetcher: sequential access across 1,000 pages produces a strong reduction in data-TLB misses, indicating that the prefetcher fills the dTLB. Skylake-X’s TLB structure is described in detail, including its separate instruction and data L1 TLBs and unified second-level TLB. The finding does not show whether the prefetcher also fills the second-level TLB or the instruction TLB. To investigate its trigger, the author proposes comparing accesses to the first and second halves of each page, then probing for prefetched translations. The supplied material begins this experiment but does not include results establishing the trigger threshold, prefetch distance, or direction.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.