DynaRL: Flexible and Dynamic Scheduling of Large-Scale Reinforcement Learning Training
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15·
summary
DynaRL claims a system that reclaims idle GPUs and reallocates them online using a dynamic hypergraph + a unified interrupt and migrate APIs, with three migration strategies (reboot, workload, peer‑to‑peer) and a context‑aware data router to preserve KV cache locality. It runs on large H100 clusters with 1.5B/7B/32B models for math‑reasoning and agentic RL, and reports up to 1.98× end-to-end speedup with negligible scheduling overhead.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 3:09Static allocation challenges
- 6:02Design principles
- 7:31Dynamic hypergraph
- 8:58Migration interface
- 11:50Architecture overview
- 12:52Scaling algorithm
- 16:31Request-level scheduling
- 17:36Evaluation results
- 18:17Conclusion
related talks
17mScaling Reinforcement Learning
17mReinforcement Learning with Human Feedback in Distributed Systems
36mAccelerating AI training fleets with sched_ext
20mEfficient Reinforcement Learning
22mAWARE: Automate Workload Autoscaling with Reinforcement Learning in Production...
16mDistRS: Disaggregated Reward Service for RLVR with Batch-Level Constraint
26mMSRL: Distributed Reinforcement Learning with Dataflow Fragments
37mA practical perspective on training reinforcement learning bots.