Paper Reading｜HM-ANN When ANNS Meets Heterogeneous Memory

This post analyzes the HM-ANN algorithm, a novel approach for Approximate Nearest Neighbor (ANN) search that addresses memory heterogeneity by optimizing index placement across DRAM and Persistent Memory (PMem). It details HM-ANN's graph construction (top-down insertion, bottom-up promotion with high-degree node selection) and search algorithms (fast memory search, parallel layer-0 search with prefetching). The evaluation compares HM-ANN against other ANN algorithms like HNSW and Disk-ANN on billion-scale datasets, highlighting its performance benefits in terms of latency and recall, especially when datasets exceed DRAM capacity. It also discusses the effectiveness of high-degree promotion and the performance gains from memory management techniques like data prefetching.