Ray: A System for High-performance, Distributed Python Applications
summary
Ray claims to scale Python from a laptop to a cluster by making functions into remote tasks and actors scheduled across nodes with a global control store and zero-copy object reads, and it exposes libraries like Tune and RLlib for hyperparameter search and reinforcement learning. The talk cites neural nets growing ~35× every 18 months, describes an open-source system spun out as AnyScale, and notes production deployments and resource-aware scheduling that actually work for ML workloads.
AI-written summary · about
chapters (28)
- 0:00Start
- 1:31The ML Landscape Today
- 2:19The Ray Vision: Sharing a common Framework
- 3:03API - Designed to Be Intuitive and concise
- 8:20How Does This Work?
- 10:59Community and Resources
- 11:53If you're already using...
- 12:38Ray Libraries
- 12:56Reinforcement Learning - Ray RLlib
- 13:49Go as a Reinforcement Learning Problem
- 14:11Growing Number of RL Applications
- 14:49RLlib: A Scalable, Unified Library for RL
- 15:25Broad Range of Scalable Algorithms
- 15:41Diverse Compute Requirements Motivated Creation of Ray!
- 17:00Hyperparameter Tuning - Ray Tune
- 18:19Nontrivial Example - Neural Networks
- 18:52Hyperparameters Are Important for Performance
- 19:05Why We Need a Framework for Tuning Hyperparameters
- 19:22Tuning + Distributed Training
- 19:46Native Integration with Tensor Board HParams
- 20:58What Are Microservices?
- 21:28Conway's Law - Embraced
- 22:17Separate Responsibilities
- 22:34Separate Management
- 23:44Management - Simplified
- 24:08What about Kubernetes (and others...)?
- 24:32Conclusion
- 25:01About Anyscale, Inc
related talks
24mRay: A System for High-performance, Distributed ML Applications
15mRay: A System for Distributed Applications
42mRay: A distributed system for emerging AI applications
01h 30mDistributed Python with Ray-Hands on with the Ray 2.0 APIs for scaling Python Workloads
01h 31mIntroduction to Ray for distributed and machine learning applications in Python
57mIntroduction to Ray for distributed and machine learning applications in Python
22mRay, a Unified Distributed Framework for the Modern AI Stack
31mScaling Python to thousands of nodes with Ray