Completed
Minimizing Effects of Asynchrony on Convergence
Class Central Classrooms beta
YouTube videos curated by Class Central.
Classroom Contents
Dorylus - Affordable, Scalable, and Accurate GNN Training with Distributed CPU Servers and Serverless Threads
Automatically move to the next video in the Classroom when playback concludes
- 1 Intro
- 2 Machine Learning
- 3 Graph Neural Networks
- 4 Stages of a Graph Neural Network
- 5 GPUs Are Not a Good Fit for Graph Operations
- 6 Combining CPUs and GPUs is Cost-Ineffective
- 7 Using Many CPU Servers Can Still Be Expensive
- 8 Key Insight: Serverless Fits Our Goals
- 9 Serverless Achieves Low-Cost, Scalable Efficiency
- 10 Challenges with Using Serverless
- 11 Challenge 1: Limited Resources
- 12 Solution: Computation Separation
- 13 Dorylus Architecture
- 14 Flow of Decomposed Tasks
- 15 Challenge 2: Limited Network
- 16 Solution: Create Pipeline of Decomposed Tasks
- 17 Data Chunks Moving Through Layer of Pipeline
- 18 Synchronize after Scatter Hinders Pipeline
- 19 Two Sync Points Makes Asynchrony Difficult
- 20 Minimizing Effects of Asynchrony on Convergence
- 21 Serverless Optimizations
- 22 Data Graphs
- 23 We Evaluated Several Aspects of Dorylus
- 24 High Value on Large-Sparse Graphs
- 25 Dorylus Outperforms Existing Systems
- 26 Dorylus Scales Full Graph Training
- 27 Conclusion: Dorylus Provides Value
