2. RISECamp Overview
• Day 1: Systems support for emerging AI applications
• RL Concepts
• Ray: a distributed exec. framework for emerging AI apps
• Clipper: a low latency prediction serving system
• Integration: Ray-trained policy served on Clipper
• Day 2: Data Analytics and Security
• PyWren: scalable data analytics with AWS Lambdas
• Ground : contextual data analytics
• WAVE: Global-scale authorization for IoT without central
authority
2
3. Ray
Ease of use
• Minimal changes to parallelize existing Python serial code
Performance
• Millions of tasks per second with msec level latencies
• Adapt to changing environments in real-time
Flexibility
• Combine neural networks, planning, search, simulation, etc
• Heterogeneous tasks: CPUs/GPUs, durations, computation
• Fine-grained data and task dependencies, dynamic execution
4. Ray Tutorial
• A set of exercises designed to demonstrate
• Ray API
• how to parallelize existing code with Ray
• how to parallelize common ML training pipelines on Ray
• How to encapsulate mutable state with actors
• Ray Reinforcement Learning Library (Rllib)
5. Ray Status and Roadmap
• Status
• Released 0.2.0, working towards 0.2.1
• Main 0.2.0 addition: a collection of RL algorithms on Ray
• Improved object store interaction and performance
• Roadmap
• Improved actor fault tolerance
• Improved scheduling policies
• Scalability, stability, and performance
6. Clipper
Predict
MC MC
RPC RPC RPC RPC
Clipper: Decouples Applications and Models
Applications
Model Container (MC) MC
Caffe
7. Clipper Tutorial
• Develop conceptual model of how Clipper works
• Deploy library and user-defined ML models to Clipper
• Send queries to Clipper-served models
• Use Clipper as a tool to power your own ML applications
8. Clipper Status and Roadmap
• Status:
• Release 0.2.0 coming out soon!
• Single command deployment in production environments
• Support for Python, Java, R ML frameworks
• Support for Kubernetes
• Roadmap:
• System performance and debuggability
• 5x performance improvement on serving latency
• Model debuggability before and after deployment
10. PyWren Tutorial
• Simple machine learning tasks with PyWren + lambdas
• PyWren API
• Advanced exercises using PyWren for :
• Large matrix computations
• Hyperparameter optimization
11. PyWren Status and Roadmap
• Status:
• 0.3 released
• Improved serialization and scaling (up to 3000 lambdas)
• Roadmap:
• LambdaPack for large-scale matrix computation
• Support for GPUs in standalone mode
• Increased performance compatibility with Spark
• Support for jobs exceeding lambda limits
12. PyWren Status and Roadmap
• Status:
• 0.3 released
• Improved serialization and scaling (up to 3000 lambdas)
• Roadmap:
• LambdaPack for large-scale matrix computation
• Support for GPUs in standalone mode
• Increased performance compatibility with Spark
• Support for jobs exceeding lambda limits
13. Ground
• Tutorial:
• Familiarity with Ground and how to use it
• Integrating data context from multiple sources in a single system
• Status:
• Released 0.1.2, solid core system, looking for more use cases
• Roadmap:
• Exploring layers above & below Ground
15. WAVE
• Tutorial:
• Familiarity with the system and its API
• Ability to delegate authority without central coordination
• Status:
• 2.2.0 released
• Full decentralization
• Consistent routing object views
• Ether currency and smart contracts
16. Recap: Focus of 1st RISE Camp
Reinforcement Learning (RL): act in dynamic environments
Prediction serving: make accurate and timely decisions at scale
Ease of development: bring ML to non-CS researchers
Data context service: capture context in which data is used
Security: decentralized authorization for IoT