Elastic Resource Sharing for Distributed Deep Learning
summary
Proposes Apathetic Future Share (AFS), an elastic resource‑sharing algorithm that balances short‑job prioritization and per‑job scaling efficiency, and shows in cluster traces that AFS cuts average job completion time versus SRTF, Themis and Tiresias‑L by up to 2.7x, 2.2x and 3.1x respectively. Implements a system called CoDDL that automatically parallelizes models, performs frequent fine‑grained GPU share re‑adjustments, and uses concurrent reconfiguration plus fast canceling to curb thrashing; CoDDL materially reduces reconfiguration‑induced slowdown (≈2.7× in their replay) and realizes AFS’s gains.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:31Overview
- 3:38Algorithm Size Challenge
- 5:28Resource Efficiency
- 5:59AFS
- 6:48AFSL
- 7:44AFSP
- 8:22Challenges
- 9:47KODO
- 10:58Summary
related talks
24mTiresias: A GPU Cluster Manager for Distributed Deep Learning
30mGPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
19mResource Management Innovation for AI Workloads in Cloud Computing
22mThemis: Fair and Efficient GPU Cluster Scheduling
09mDemocratizing Access to GPUs for Faster MLAI Innovation with Generative AI
19mEfficient LLM Serving on Commodity GPU Clusters with Data-Reduced Cross-Instance...
20mScheduler and Sharding Considerations for Network Efficiency
22mResearch SuperCluster(RSC): Accelerating AI Research at Scale