Splitting complex model inference between multiple computing devices can overcome latency and energy constraints at the edge. Newer edge accelerator devices with higher computing capacity and energy efficiency, enable more fine-grained offload throughout layers of the network, leading to the potential for multiple split configurations. However, optimizing a DNN for inference across networked devices requires a precise performance model that can guide design choices. In this paper, we demonstrate that existing models for computation and communication latency are inaccurate due to system considerations and propose a new empirical model based on structured benchmarking, considering data ingestion overhead due to transfers between devices as well as within a device for data to reach the GPU. We validate our split inference performance model using VGG16 and ResNet50 networks on two heterogeneous platforms, showing it achieves a mean absolute error of no more than 4% for both DNNs, significantly outperforming previous models with errors of more than 15%. We also validate the practical utility of our model by incorporating it into existing split inference search algorithms under multi-split, dynamic bandwidth, and multi-tenant scenarios, demonstrating its effectiveness in navigating the split inference search space.
Hao Liu, M. Fouda, Ahmed M. Eltawil et al.· International Conference on...· 0 citations
Integrating uncrewed aerial vehicle (UAV)-mounted aerial reconfigurable intelligent surfaces (RISs) holds significant promise for enhancing the performance of ground-based networks. This paper proposes a novel three-dimensional (3D) deployment and partitioning framework for aerial RIS-assisted uplink grant-free non-orthogonal multiple access (GF-NOMA). GF-NOMA allows users to access the resource block immediately without scheduling overhead, but requires sufficient received power disparity for reliable successive interference cancellation (SIC). Specifically, we consider three design objectives, namely max-sum throughput (MST), max-min fairness (MMF), and proportional-fairness rate (PFR), and jointly optimize aerial RIS partitioning and deployment. A closed-form analytical solution is derived for MST and MMF regimes, while an unsupervised learning (USL) framework is developed for partitioning, and a deep reinforcement learning (DRL)-based policy is designed for adaptive UAV deployment under imperfect channel estimates and residual SIC conditions. Extensive numerical results show that the learned schemes for MST and MMF closely track their respective theoretical benchmarks, within 1% for MST and $2-10\%$ for MMF, under the same non-ideal channel knowledge model. All three proposed USL-DRL schemes significantly outperform fixed-deployment baselines: the proposed MST-USL-DRL achieves approximately 32–34% sum-rate gain, the proposed MMF-USL-DRL improves the minimum user rate by about 47–74% depending on the evaluation regime, and the proposed PFR-USL-DRL delivers roughly 57% gain over fixed deployment.
Tomiris Altay, Mohd Hamza Naim Shaikh, A. Çelik et al.· IEEE Transactions on Cogniti...· 0 citations