Lune

ICDE2026顶会

Reconfiguring Scalable Hashing with Persistent CPU Caches

Zhenyu Yu, Bolong Zheng, Ling Xu, Qianlu Wu, Qiang Chen, Ziyang Yue

2026年份

摘要

As an essential functionality in main memory system, the extendible hash index exhibits both low latency and instant recovery with the persistent memory (PM). The recently released eADR feature of PM further provides an opportunity to enhance write performance of extendible hash index. Compared to the previous generation ADR, eADR avoids inefficient manual flush and mitigates write amplification by including CPU cache into the persistent domain. However, existing PM-based hash indexes are either designed without the enhancement of eADR, or fail to effectively leverage the persistent CPU cache. On the one hand, although using DRAM as a buffer helps mitigate write amplification, its shared usage of the CPU cache with PM leads to severe cache contention. On the other hand, skewed access towards metadata leads to inefficient CPU cache utilization. We propose PIONEER, an eADR-friendly persistent extendible hash index for a hybrid DRAM-PM memory architecture. First, to alleviate cache contention during data persistence, PIONEER introduces the directory snapshot that employs non-temporal write to bypass the CPU cache while preserving crash consistency guarantees. Second, PIONEER enhances the CPU cache utilization by re-designing the layout of hash index in a metadataseparated manner. In addition, PIONEER proposes a data-aware persistence strategy that dynamically adjusts the persistence approach for each segment based on data hotness. Extensive experiments conducted with YCSB workloads demonstrate that PIONEER achieves up to 36.4×36.4 \times higher write performance compared to other state-of-the-art persistent hash indexes.

问问这篇 Paper

问问你的智能体。

Lune 读过与它相关的顶会 Paper,每个回答都会注明依据哪几篇。

可以从这些问题问起

智能体调用

Lunesearch_papers

在 Lune 里问

免费开始,无需绑卡

相关 Paper

黄昏的海面,两侧是细线勾勒的悬崖