STEP: Spatial Footprint Prefetcher with Multi-Point Temporal Triggers

Modern processors continue to face the memory wall, which heavily impacts computer performance. Besides sophisticated cache hierarchies, prefetching is a proven technique to hide latency and mitigate the problem. Spatial footprint prefetchers have attracted research interest due to their ability to cover diverse memory access patterns with relatively low hardware cost. However, current implementations trigger on a single-point-in-time event, forcing a trade-off between issuing prefetches early, which increases noise, and issuing them later, which reduces opportunities. We propose STEP, a spatial footprint prefetcher that introduces multiple sequential temporal decision points within a page's cache lifetime and selectively issues prefetches when confidence is high. STEP coordinates these decisions through a lightweight prefetch-confidence evaluator, enabling early opportunities while retaining high accuracy with later prefetching triggers. Moreover, STEP consolidates metadata into a single Pattern History Table (PHT), reducing storage overhead. Experimental results on SPEC CPU2006, SPEC CPU2017, and CloudSuite show that STEP improves performance across a broad range of workloads. In the L2 single-core evaluation, STEP remains ahead of the strengthened ISO-storage baseline eBingo, while in the L1 single-core evaluation, STEP further outperforms Gaze, the strongest baseline at that level. STEP also delivers a stronger performance-storage trade-off, as eBingo requires substantially larger metadata capacity to approach STEP's lowstorage operating point.