The Clustering Strikes Back: Building Cost-Effective and High-Performance ANNS at Scale
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15·
summary
Helmsman replaces memory-heavy HNSW with a clustering-based ANNS that uses SPDK userspace storage, a leveling-learned pruning model to pick how many clusters to probe per query, and GPU-accelerated clustering plus elastic CPU splitting for fast rebuilds. It targets production needs: 5–10 ms latency at million-QPS scale, top-K in the hundreds to thousands, and over 300 billion vectors. Production deployment reduces footprint from ~35,000 CPU cores and 0.35 PB DRAM to 40 machines, claims >90% hardware cost savings, and enables billion-scale index rebuilds within hours.
AI-written summary · about
chapters (9)
- 0:00High-performance ANN search
- 1:00Production requirements
- 4:33Graph indexes challenges
- 7:05Clustering for SSD efficiency
- 9:12Hermes design
- 11:40GPU index construction
- 12:24Evaluation results
- 13:33Cost efficiency at Renren
- 14:59Key takeaways
related talks
16mTowards High-throughput and Low-latency Billion-scale Vector Search via CPU/GPU...
18mScalable Billion-point Approximate Nearest Neighbor Search Using SmartSSDs
22mCXL-ANNS: Software-Hardware Collaborative Memory Disaggregation and Computation...
01h 09mVector Search: Ask Me Anything!
45mBehind the hype: managing billion-scale embeddings in Elasticsearch and OpenSearch
32mRetrieval Augmentation and Semantic Search at Scale
47mCapacity Planning, and Scaling/Optimization for Vector Workloads
21mDesigning Scalable Networks for Large AI Clusters: Challenges and Key Insights