Towards High-throughput and Low-latency Billion-scale Vector Search via CPU/GPU...
summary
FusionANNS is a working SSD + single entry‑level GPU ANNS that uses CPU/GPU collaborative filtering to keep PQ‑compressed vectors in GPU memory, avoid swapping, and limit SSD reads. It implements multi-tiered indexing, heuristic re-ranking, and redundant-aware I/O deduplication to cut I/O and computation; evaluation against SPANN and RUMMY reports 9.4–13.1× higher QPS (5.7–8.8× cost efficiency) vs SPANN and 2–4.9× QPS (2.3–6.8× cost efficiency) vs RUMMY while preserving low latency and accuracy.
AI-written summary · about
chapters (5)
- 0:00Background and motivation
- 2:27Performance challenges
- 7:47FusionANNS system design
- 13:33Evaluation and results
- 15:47Conclusion
related talks
18mScalable Billion-point Approximate Nearest Neighbor Search Using SmartSSDs
16mThe Clustering Strikes Back: Building Cost-Effective and High-Performance ANNS at Scale
12mAchieving Low-Latency Graph-Based Vector Search via Aligning Best-First Search...
15mFast Vector Query Processing for Large Datasets Beyond GPU Memory with Reordered...
32mRetrieval Augmentation and Semantic Search at Scale
40mUnder the hood of vector search with JVector
19mHow Microsoft powers planet-scale AI apps with DiskANN
35mBuilding High Performance Search Indexes in Rust with SIMD