Software-hardware collaborative CXL memory dissaggregation system for billion-scale approximate nearest neighbor search

Abstract

In this thesis, we propose software-hardware collaborative compute express link (CXL) memory disaggregation system for billion-scale approximate nearest neighbor search (ANNS). ANNS is widely used in commercial services such as image search, database, and recommendation systems by reason of high accuracy and low latency. However, in production-level ANNS, there is a challenge of requiring a large amount of memory due to the extensive dataset. CXL-ANNS is composed CXL memory dissaggregation system that can accomodate the billion-point graph dataset. Additionally, by using near data processing and data prefetch, the performance of CXL-ANNS is improved. Proposed CXL-ANNS exhibits $93.3%$ lower latency than state-of-the-art ANNS platforms.