主讲人
Hengrui Wang
Tsinghua University
时间
2026年3月17日 星期二
下午 14:00-15:00
地点
上财计算机与人工智能学院104会议室
Abstract
With the increased bandwidth and parallelism of modern storage media such as NVMe SSDs, together with the reduced per-I/O latency, the bottlenecks of LSM-Tree key-value stores have gradually shifted from pure I/O amplification to CPU overhead, resource contention between foreground queries and background compaction, and cache efficiency for range scans. As a result, a large body of prior work that focuses solely on balancing read amplification and write amplification in LSM-Trees becomes less applicable in this new regime. My work aims to maximize read performance without sacrificing LSM-Tree write performance with three complementary techniques.
Biography
Hengrui Wang is a PhD student in the Institute for Interdisciplinary Information Science (Yao Class) at Tsinghua University, advised by Prof. Huanchen Zhang. Before joining Tsinghua, he obtained his BS at Yuanpei College of Peking University, advised by Prof. Tong Yang. His research interests mainly lie in designing lightning-fast data structures and algorithms for storage, databases, and machine learning systems.
