Tiresias: A GPU Cluster Manager for Distributed Deep Learning
summary
Tiresias is a GPU cluster manager that actually implements two age‑based schedulers — Discretized Two‑Dimensional Gittins index (needs a prior distribution) and Discretized Two‑Dimensional LAS (needs no prior) — plus a placement policy driven by a model profiler that measures tensor‑size skew to decide when consolidation matters. A prototype and trace‑driven sims show up to 5.5× average job‑completion‑time improvement over an Apache YARN baseline, comparable to clairvoyant SRTF and roughly 2× better than Gandiva.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:32GPU Cluster for Deep Learning Training
- 1:43GPU Cluster Manager
- 3:01Challenge I: Unpredictable Training Time
- 4:30Challenge II: Over-Aggressive Job Consolidation
- 7:58Characteristics of DL Training Jobs
- 9:45Age-Based Schedulers
- 11:07Two-Dimensional Age-Based Scheduler (2DAS)
- 12:072D-Gittins Index Partial Information
- 15:34Characteristics of DL Models
- 17:13Model Profile-Based Placement
related talks
30mGPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
12mEasy and affordable access to GPUs for AI/ML workloads
12mElastic Resource Sharing for Distributed Deep Learning
20mScheduler and Sharding Considerations for Network Efficiency
24mImproving GPU Utilization and Accelerating Model Training with Scheduling Framework and NRI
31mAI, CERN, and the Quest for GPU Custody: How CERN Leverages DRA for Efficient GPU Sharing
37mDistributed Deep Learning on Apache Mesos with GPUs and Gang Scheduling
19mAnalysis of Large-Scale Multi-Tenant GPU Clusters for DNN Training Workloads