Application of Deep Reinforcement Learning in Optimizing Virtualized GPU Resource Scheduling for Real-time Rendering

A closed-loop DRL scheduler for virtualized GPUs under time-varying rendering workloads.

Investigating deep reinforcement learning for virtualized GPU scheduling, addressing resource underutilization and latency instability caused by time-varying real-time rendering workloads; formulating adaptive resource allocation as a sequential decision-making problem.
Developing a closed-loop scheduler that converts VM/GPU telemetry into workload states and dynamically selects GPU partitioning, preemption, and priority actions, with a reward function balancing utilization, task latency, service quality, and reconfiguration overhead.
Evaluating the learned policy against static partitioning and fixed-priority baselines under changing execution conditions using GPU utilization, tail latency, SLA violations, and scheduling overhead; preparing an EI-indexed journal manuscript on the resulting framework and experimental findings.

Tech stack