Ray Data on GPU - Breaking Preprocessing Bottleneck for Model Training
summary
Stripe moves large tabular pipelines (1.9T annual volume, 6–10 TB / ~2B rows) onto GPUs by adding cuDF as a Ray Data batch format, using Ray map_batches plus a GPU hash-shuffle (Rapids) and grouped-aggregation primitives, plus heterogeneous pipelining and intermediate map-reduce steps to trim long-tail category payloads. Synthetic runs show ~5x speedup and ~82% cost reduction; production fit shows ~32x speedup with ~92% cheaper hardware. The GPU-accelerated Ray Data preprocessors are implemented and yield stable, reproducible gains. cuDF on GPU map_batches + GPU shuffle
AI-written summary · about
chapters (7)
- 0:00Introduction and overview
- 4:42Preprocessor level optimizations
- 8:35Optimizing preprocessor chains
- 10:19Results on synthetic data
- 12:01GPU optimizations by Nvidia
- 21:11Lessons learned with GPUs
- 22:58Q&A session
related talks
30mHow to Scale Unstructured Data Processing with Ray Data
32mAnyscale's Ray Data: Revolutionizing Batch Inference
22mLarge Scale Data Loading and Data Preprocessing with Ray
33mOvercoming Distributed ML Challenges with Ray Train
01h 08mGPU Data Science with Rapids
41mAccelerating Data Science with RAPIDS
32mRay @ eBay: Pioneering a Next-Gen AI Platform
27mPinterest's ML Evolution: Distributed Training with Ray