SLIDE 1
ADMINISTRIVIA
- Assignment 1 Grades
- Assignment 2 due on Fri
- Course Project emails
CS 744: RAY Shivaram Venkataraman Fall 2019 ADMINISTRIVIA - - - PowerPoint PPT Presentation
CS 744: RAY Shivaram Venkataraman Fall 2019 ADMINISTRIVIA - - - PowerPoint PPT Presentation
CS 744: RAY Shivaram Venkataraman Fall 2019 ADMINISTRIVIA - Assignment 1 Grades - Assignment 2 due on Fri - Course Project emails Bismarck Supervised learning, Unified Interface Shared memory, Model fits in memory Parameter Server Large
SLIDE 2
SLIDE 3
Machine Learning Bismarck Supervised learning, Unified Interface Shared memory, Model fits in memory Parameter Server Large datasets, large models (PB scale) Consistency model, Fault tolerance Tensorflow Need for flexible programming model Dataflow graph Heterogeneous accelerators
SLIDE 4
Bismarck Parameter Server Tensorflow
WORKLOADS
SLIDE 5
REINFORCEMENT LEARNING
SLIDE 6
SLIDE 7
SLIDE 8
RL SETUP
SLIDE 9
RL REQUIREMENTS
Simulation Training Serving
SLIDE 10
RAY API
Tasks Actors
futures = f.remote(args) actor = Class.remote(args) futures = actor.method.remote(args)
- bjects = ray.get(futures)
ready = ray.wait(futures, k,timeout)
SLIDE 11
COMPUTATION MODEL
SLIDE 12
ARCHITECTURE
SLIDE 13
Global control store
Object table Task table Function table
SLIDE 14
RAY SCHEDULER
Global Scheduler Global Control Store
SLIDE 15
FAULT TOLERANCE
Tasks Actors GCS Scheduler
SLIDE 16
DISCUSSION
https://forms.gle/QQyLbwjAufJNXWnr6
SLIDE 17
Consider you are implementing two task: a deep learning model training and a sorting application. When will use tasks vs actors and why ?
SLIDE 18
SLIDE 19
Considering AllReduce using MPI as the baseline parallel programming task. Discuss the improvements made by MapReduce, Spark over MPI and discuss if/how Ray further contributes to the comparison.
SLIDE 20