PathWeaver: A High-Throughput Multi-GPU System for Graph-Based Approximate...
summary
PathWeaver is a multi-GPU ANNS system that implements pipeline-based path extension and ghost staging to pass local search results between GPUs and a tiny representative graph to bootstrap the pipeline, plus direction-guided selection to filter neighbors with a 1-bit edge compression before full L2 checks. Evaluation uses four RTX 6000 GPUs connected by ambulink bridge on six datasets (1M–50M vectors, dims 96–960) and reports 3.24× geomean speedup and up to 5.30× at 95% recall.
AI-written summary · about
chapters (9)
- 0:00PathWeaver introduction
- 1:43Graph-based approximate search background
- 3:47Multi-GPU ANS scalability needs
- 5:14Pipeline pass extension for GPUs
- 9:05GO staging for first stage
- 11:13Direction-guided neighbor selection
- 13:43Optimizations' combined impact
- 14:46Performance evaluation results
- 15:50Conclusion and open-source release
related talks
25mPathfinder to GPU Offload in WASM
45mHigh-Performance and Scalable Middleware for HPC and Deep...
16mTowards High-throughput and Low-latency Billion-scale Vector Search via CPU/GPU...
16mRevisiting Pipeline Parallelism for LLM Serving
20mHetPipe: Enabling Large DNN Training on (Whimpy) Heterogeneous GPU Clusters through
13mGNNAdvisor: An Adaptive and Efficient Runtime System for GNN Acceleration on GPUs
47mDesigning High-Performance Scalable Middleware for HPC, AI, & Data Science in Exascale Systems & ...
26mPetabyte-Scale Video Curation on CoreWeave