利用儲存內計算加速向量資料庫搜尋

Abstract

向量資料庫（Vector Databases, VDBs）透過向量搜尋實現高效的資訊檢索，在現代資料處理中至關重要。 雖然近似最近鄰搜尋（Approximate Nearest Neighbor Search, ANNS）方法（如基於圖的和基於叢集的方法）解決了大型資料集在計算上的挑戰。然而，大量資料移動所導致的效能瓶頸，限制了向量搜尋的可擴展性和效率。 透過儲存內計算 (In-storage Computing) 技術能夠有效緩解此瓶頸所造成的性能損失。 然而，過去的研究缺乏存儲內計算裝置內部的量化分析，既無法直觀的說明效能提升的原因，系統中也可能存在潛在的效能瓶頸。 本研究提出了一種基於 COSMOS+ OpenSSD 平台的儲存內向量搜尋裝置設計與實現。借助實際平台的分析發現直觀的設計會導致指令管線化效能下降並針對此問題提出改善方案。我們的方法在距離計算功能的效上相較於現有的儲存內計算解決方案提升了 1.49 倍，達到了理想效能的 88%。此外，我們的方案在端到端吞吐量上，與儲存內計算方法相比，提升了 1.42 ~ 1.44 倍; 與使用傳統 SSD 方法相比，提升了 1.75 ~ 2.03 倍。最後，我們分析了引入儲存內計算所帶來的新的系統瓶頸，並且針對此瓶頸提出未來的研究建議。

Vector Databases (VDBs) enable efficient information retrieval through vector search, crucial in modern data processing. While Approximate Nearest Neighbor Search (ANNS) methods address computational challenges, data movement bottlenecks limit the scalability and efficiency of vector search. Prior In-storage computing (ISC) reserch offers a solution but lacks quantitative analysis on real devices, making it difficult to explain performance gains and identify potential bottlenecks. This study presents the design and implementation of an in-storage vector search device based on the COSMOS+ OpenSSD platform. Through real-device analysis, we identify a key bottleneck where a naive PE design leads to pipeline inefficiencies between commands. To address this, we propose an optimized approach that improves distance computation performance by 1.49x, achieving 88% of the theoretical peak performance. Additionally, our design enhances end-to-end throughput by 1.42x ~ 1.44x compared to existing ISC solutions and by 1.75x ~ 2.03x over traditional SSD-based methods. Finally, we analyze new system bottlenecks introduced by ISC and provide insights for future research directions.