Decentralized Application-Level Adaptive Scheduling for Multi-Instance DNNs on...
summary
A user‑level scheduler implements decentralized DQN agents (lightweight 2–5 fully connected layers) inside each app to pick execution configurations (CPU threads, GPU/DSP/codegen, TF‑Lite modes), measures latency and power via Android battery tools, and adapts to different OS versions and uncontrollable apps like TikTok/YouTube. The system empirically converges to a Nash equilibrium (agents stabilize after hundreds to a few thousand iterations, ~6,600 in one trace) and yields consistent speedups and energy savings on a Galaxy S21 and a tablet.
AI-written summary · about
chapters (14)
- 0:00Decentralized DNN scheduling
- 1:31Mobile multi-instance DNN challenges
- 4:22Configurable execution options
- 7:17Problem definition and constraints
- 9:45Research questions on adaptive scheduling
- 12:13Latency vs power trade-offs
- 13:40OS interaction and design
- 15:05Nash equilibrium convergence
- 16:49Experimental setup and hardware
- 18:14Performance vs baseline
- 20:27Learning speed and reward analysis
- 21:32Stability over time
- 21:53Handling background apps
- 22:36Summary and conclusion
related talks
15mEffectively Scheduling Computational Graphs of Deep Neural Networks toward Their...
26mNeuOS: A Latency-Predictable Multi-Dimensional Optimization Framework for DNN-drive
19mDynaRL: Flexible and Dynamic Scheduling of Large-Scale Reinforcement Learning Training
17mUnleash All Cores: Asymmetry-Aware Scalable DNN Inference on Mobile CPUs
36mAccelerating AI training fleets with sched_ext
38mNEURAL NETWORKS ON YOUR MOBILE PHONE. DIVING INTO THE NN API
17mModel-Switching: Dealing with Fluctuating Workloads in Machine-Learning-as-a-Service
10mFair Scheduling for Deep Learning Workloads in Kubernetes