Top-k nearest neighbor (kNN) search based on vector similarity is a cornerstone in numerous domains, including AI applications such as retrieval-augmented generation. When datasets exceed the capacity of main memory, they must be stored on storage devices (e.g., flash drives), necessitating efficient near-storage search mechanisms. In this talk, we present two distinct near-storage systems tailored for both exact and approximate kNN queries.
ES4D is an exact vector similarity search platform that executes queries directly on solid-state drives (SSDs). It accelerates kNN search through two levels of early termination, incorporating pre-clustering and vector sharding, as well as additional optimizations. Offloading kNN search to the SSD not only reduces energy consumption but also enables ES4D to optimize the physical page placement of dataset vectors, thereby enhancing throughput. Our evaluations on various vector datasets demonstrate that ES4D achieves a 1.9× improvement in search performance compared to existing exact kNN solutions while reducing energy consumption by 8.1×.
GraphAccel is a near-storage accelerator designed for efficient, graph-based approximate vector similarity search. It adopts an optimized page packing mechanism to minimize SSD page accesses per query, coupled with a speculative search strategy that fully utilizes idle SSD chips and channels. These innovations deliver significant performance gains over SSD-based graph search solutions such as DiskANN and DiskANN++.

Hyungmin Cho received the B.S. degree in computer science engineering from Seoul National University, South Korea, in 2005, and the M.S. and Ph.D. degrees in electrical engineering from Stanford University, in 2010 and 2015, respectively. He was a Research Scientist with Intel Labs, Santa Clara, CA, USA. From March 2017 to August 2019, he was an Associate Professor with Hongik University, South Korea. He is currently an Associate Professor with the Department of Computer Science and Engineering, Sungkyunkwan University (SKKU), South Korea. His research interests include reliable computer systems and accelerator architectures for high-performance computing.