Ray: A System for Distributed Applications
summary
Ray scales Python from laptop to cluster by turning functions into ray.remote tasks and objects into actors for stateful servers. Demo: Monte Carlo π runs on one core (≈10 s for 1M samples) but uses 4 cores for ~3–4× speedup when Ray starts 16 workers. A parameter‑server built as an actor with 5 trainers (10k params, 500 iterations) drops wall time from ~6.3 s user time to 37 ms user time when distributed. Also ships Tune hyperparameter search for parallel model tuning.
AI-written summary · about
chapters (8)
- 0:00Start
- 11:07The Ray Vision: Sharing a common Framework
- 11:39Hyperparameter Tuning - Ray Tune
- 12:21Nontrivial Example - Neural Networks
- 12:48Why We Need a Framework for Tuning Hyperparameters
- 13:00Tuning + Distributed Training
- 13:20Separate Management
- 13:40Management - Simplified
related talks
24mRay: A System for High-performance, Distributed ML Applications
26mRay: A System for High-performance, Distributed Python Applications
01h 31mIntroduction to Ray for distributed and machine learning applications in Python
42mRay: A distributed system for emerging AI applications
22mRay, a Unified Distributed Framework for the Modern AI Stack
11mKeynote Q&A
24mDistributed Black-Box Model Explanation with Ray
24mBuilding Complex Data Analytics Pipelines with Ray