AOEH: An Efficient Extendable Hashing to Reduce Read/Write Amplification for Persistent Memory

With the exponential growth of data volumes and index sizes, DRAM capacity has become a critical bottleneck in modern high-performance systems. Expanded memory architectures, particularly those leveraging Persistent Memory (PM), offer a promising solution by combining high capacity, byteaddressability, and non-volatile characteristics with near-DRAM performance. Although significant efforts have been made to adapt hashing schemes for PM, existing PM-optimized hash indexes still suffer from substantial read and write amplification, severely degrading overall performance. To address that, this paper proposes AOEH, an Efficient Extendible Hashing scheme to Reduce Read/Write Amplification for Persistent Memory. Based on our systematic PM I/O amplification analysis, AOEH integrates a principled EH-based structure with a compact, self-adapting fingerprint to eliminate key dereferencing and reduce read amplification, complemented by a one-way lock-free mechanism with deterministic semantics to tackle write amplification. Extensive evaluation on YCSB workloads shows that AOEH delivers superior throughput across diverse workloads, reaching an average of $3.62 \times$ and $3.28 \times$ the throughput of the baseline designs in read-intensive and writeintensive scenarios, respectively, while reducing read and write amplification by 49% and 21% over the best-performing baseline.