Scalable Billion-point Approximate Nearest Neighbor Search Using SmartSSDs
summary
SmartANNS implements hierarchical indices on host plus Samsung SmartSSDs with FPGA acceleration, dynamic task scheduling and a gradient-boosted shard pruning model (1M training vectors, lr=0.05, 500 iters) to cut on-SSD work. On a 1B-vector dataset it reports up to 10.7× QPS versus CSDANNS and near-linear scaling as SmartSSDs are added. The system moves shard centroids to host, schedules queries to maximize reuse and balance hardness, and runs graph search kernels inside SmartSSDs for reduced IO and latency.
AI-written summary · about
chapters (13)
- 0:00SmartSSD-based ANN search intro
- 2:23Large-scale vector search challenges
- 3:27SmartSSD opportunities for ANN
- 4:31Prior SmartSSD approach limits
- 7:21SmartNS architecture design
- 9:11Shot number optimizer
- 10:17Data layout for load balancing
- 12:47Dynamic task scheduling
- 14:33Search engine in SmartSSDs
- 15:57Query workflow in SmartNS
- 16:20Performance results setup
- 17:50Comparison with GPU/SSD solutions
- 18:10Summary
related talks
16mTowards High-throughput and Low-latency Billion-scale Vector Search via CPU/GPU...
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...
01h 09mVector Search: Ask Me Anything!
32mRetrieval Augmentation and Semantic Search at Scale
13mHardware/Software Co-Programmable Framework for Computational SSDs to Accelerate Deep...
16mRIANN: Real-time Incremental Learning with Approximate Nearest Neighbor on Mobile Devices
16mCognitive SSD: A Deep Learning Engine for In-Storage Data Retrieval