Regular Paper Accepted at the Outstanding International Conference International Conference on High Performance Computing, Data, and Analytics (HiPC) 2025

The paper 'CALL: Context-Aware Low-Latency Retrieval in Disk-Based Vector Databases', written by Jung Yeon-woo (first author), Cho Hyun-ji (master's) and Park Gyu-ri (doctoral) of the Data-Centric Computing and AI Systems Laboratory (DISCOS) with Professor Kim Young-jae and Professor Park Sung-yong (corresponding author), has been accepted for publication at the 32nd International Conference on High Performance Computing, Data, and Analytics (HiPC 2025).

Disk-based vector search — storing large vector indexes on high-speed storage devices and dynamically loading into memory only the index clusters needed during approximate nearest neighbor (ANN) search — has recently attracted attention. Vector databases cache index clusters in memory to reduce disk I/O during ANN search. Because vector search queries show patterns of searching the same index clusters, however, caching strategies that do not take this into account are limited in maximizing cache utilization.

The team analysed the performance limits of the index cluster caching policies used in existing vector databases and proposed CALL to maximize cache utilization. CALL computes the similarity of the index cluster sets of each query at runtime, groups queries with a high proportion of overlapping clusters and reorders the search sequence. On the basis of the distribution of index clusters across the queries in a group, it also improves the cache hit rate by asynchronously prefetching — before the query group changes — the index clusters that the first query of the next group will search.

CALL was implemented in FAISS, a representative vector search engine; in the experiments tail latency fell by more than 33% compared with existing caching techniques, and cache utilization was also maintained stably.

Jung Yeon-woo, a doctoral student, said: "There were many challenges in analysing and overcoming system-level performance bottlenecks in a vector search engine, but it is deeply meaningful to have achieved performance improvements using various techniques learned in systems programming. I hope many students will experience the appeal and satisfaction of systems software research, optimizing today's fashionable AI applications from a systems perspective to improve their performance."

HiPC is a prestigious international conference for discussing the latest research trends and technological innovation in high-performance computing and data analytics; this year it will be held in Hyderabad, India, from 17 to 20 December 2025. HiPC is also listed among the outstanding international conferences in computer science under the BK21 Plus programme (BK=1).

References:

- The 32st IEEE International Conference on High Performance Computing, Data, and Analytics (HiPC25)

- Website : https://hipc.org   

- Paper: https://discos.sogang.ac.kr/file/2025/intl_conf/HiPC_2025_Y_Jeong.pdf