Content-Aware Scheduling of a Network of DNNs on Heterogeneous Edge MPSoCs
Guardado en:
| Publicado en: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Autor principal: | |
| Publicado: |
ProQuest Dissertations & Theses
|
| Materias: | |
| Acceso en línea: | Citation/Abstract Full Text - PDF |
| Etiquetas: |
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Resumen: | The growing deployment of machine learning (ML) workloads at the edge reflects a significant shift toward decentralized computing. The diverse capabilities and constraints of edge devices demand efficient and adaptive strategies for deep neural network (DNN) inference to support real-time, data-driven applications. This need is especially critical in latency-sensitive domains such as autonomous vehicles and augmented/virtual reality (AR/VR). However, these workloads face several challenges, including hardware heterogeneity, input-dependent execution patterns, real-time processing demands, and runtime variability caused by system-level effects. Addressing these challenges is essential for maximizing resource utilization, ensuring timely and accurate inference, and enabling adaptation to runtime variability.To this end, three complementary scheduling frameworks were developed and evaluated on a Qualcomm RB5 platform: Content-Aware Mapping of Deep Neural Networks on Edge MPSoCs (CAMDNN), Elastic Multi-Tenant Execution of DNNs on Heterogeneous Edge MPSoCs (EMERALD), and Scheduling under Uncertainty with Reinforcement Learning (SURE). These are compared against the widely used Heterogeneous Earliest Finish Time (HEFT) baseline.CAMDNN is a hierarchical, content-aware scheduling framework that partitions decisions between local and global stages. The local scheduler maps each DNN model to the most efficient compute unit, while the global scheduler applies an integer linear programming (ILP)-based approach to refine mapping under system constraints. CAMDNN achieves up to 32% lower execution time compared to HEFT and outperforms CPU-only, GPU-only, and central queue baselines by 6.67×, 5.6×, and 2.17×, respectively.EMERALD extends CAMDNN with elastic scheduling, adjusting input resolution to reduce computation while meeting frame deadlines. It lowers deadline misses by up to 11× and 12.3× compared to CAMDNN and HEFT at 30 frames per second (FPS).SURE addresses runtime variability caused by dynamic voltage and frequency scaling (DVFS), thermal throttling, and workload fluctuations. It models multi-DNN scheduling as a reinforcement learning problem using a graph neural network (GNN) encoder and pointer-based policy network. It achieves up to 6.2× and 2.0× improvements over CAMDNN and HEFT.Together, these frameworks offer a unified solution for low-latency, deadline-aware, and uncertainty-resilient scheduling of multi-DNN workloads at the edge. |
|---|---|
| ISBN: | 9798290966601 |
| Fuente: | ProQuest Dissertations & Theses Global |