CASSINI: Network-Aware Job Scheduling in Machine Learning Clusters
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
CASSINI introduces a network-aware job scheduler that packs ML jobs by their periodic traffic using a geometric abstraction (circles and a unified LCM-based circle) and an optimization that discretizes rotation angles to compute per-job time shifts, then reconciles per-link constraints with an Affinity graph and sequential assignment. Implemented on a 24-server A100 testbed with 13 models, it claims up to 1.6x average and 2.5x tail job-speedup and reduces ECN marks by up to 33x.
AI-written summary · about
chapters (6)
- 0:00Introduction to DNN training
- 1:09Interleaving communication
- 5:35Geometric abstraction
- 12:35Integrating into schedulers
- 13:31Performance evaluation
- 15:22Conclusion and summary
related talks
05mScheduling at the Edge of Reason: Multi-Cluster AI & OCM
16mSLearn - A Case for Task Sampling based Learning for Cluster Job Scheduling
05mProject Lightning Talk: Scheduling AI Workload Among Multiple Clusters
22mBetter Together: Jointly Optimizing ML Collective Scheduling and Execution Planning...
16mIntelligent Resource Scheduling for Co-located Latency-critical Services: A Multi-Model
16mMAST: Global Scheduling of ML Training across Geo-Distributed Datacenters at Hyperscale
21mDesigning Scalable Networks for Large AI Clusters: Challenges and Key Insights
21mExperience Productionizing and Operating GB200 Cluster