Completed
How should we allocate heterogeneous resources?
Class Central Classrooms beta
YouTube videos curated by Class Central.
Classroom Contents
Heterogeneity-Aware Cluster Scheduling Policies for Deep Learning Workloads
Automatically move to the next video in the Classroom when playback concludes
- 1 Intro
- 2 Hardware for ML training is becoming highly specialized and heterogeneous!
- 3 How should we allocate heterogeneous resources?
- 4 Challenge 1: Heterogeneous performance
- 5 Challenge 2: Diverse scheduling objectives
- 6 Related work
- 7 Gavel: A new heterogeneity-aware cluster scheduler
- 8 Scheduling policies to be made heterogeneity-aware
- 9 Policies as optimization problems
- 10 Allocations (x) as time fractions
- 11 Effective throughput
- 12 Performance optimizations: space sharing and placement
- 13 How do we realize an optimal allocation?
- 14 Gavel's round-based scheduling
- 15 Main questions
- 16 Gavel improves objectives on a heterogeneous cluster
- 17 Gavel can enable the same heterogeneous cluster to support higher input load
- 18 Gavel can support hierarchical policies
- 19 Gavel scales to clusters with hundreds of active jobs
- 20 Conclusion
