GPGPU Scheduling Schemes to Improve Latency and Resource Utilization

Abstract

Real-time embedded systems today need GPU scheduling methods that are fast, flexible, and energy-saving while handling different types of tasks with limited resources. Current schedulers, like federated and hierarchical ones, often have problems giving real-time guarantees, adapting to changing needs, and saving power — especially on AMD GPUs that have limited Local Data Share (LDS) and High Bandwidth Memory (HBM). This paper presents a new Hybrid GPGPU Scheduler that solves these problems. It combines static, dynamic, and machine learning (ML)-based scheduling with task models that are aware of memory limits. Using Gradient Boosting Decision Trees (GBDT)/XGBoost, the scheduler predicts task features and decides if they can meet deadlines, while smartly assigning Compute Units (CUs).

Tests on a Linux platform with PoCL and ROCm show that the scheduler beats others like RTGPU, Self-Suspension, and Enhanced MPCP. It improves task scheduling by 11% at high usage (U = 2.0), accepts 85% of tasks when 12 run together, and cuts scheduling delay by 40%. It also saves 9% more energy per watt while keeping power use low (1.0–2.0 W).

Resource use stays high, with 92% CU and 85% CPU usage. These results prove that the Hybrid GPGPU Scheduler is a scalable and energy-efficient solution for real-time GPU scheduling in embedded systems, balancing performance, flexibility, and power savings in systems with tasks of different critical levels.