Graph Neural Networks Automated Design and Deployment on Device-Edge Co-Inference Systems

TL;DR

GCoDE addresses the challenge of deploying Graph Neural Networks on device-edge co-inference systems by coupling architecture search with per-operation mapping, explicitly treating device-edge communication as a dedicated GNN operation. It fuses architecture exploration with mapping decisions in a unified co-design space, guided by system-performance awareness through a GIN-based latency predictor, a LUT-based cost model, and an energy estimator. The framework uses a constraint-based random search over a one-shot GNN supernet to produce a zoo of optimal designs and deploys them via a pipelined co-inference engine and runtime dispatcher. Empirical results show dramatic improvements over state-of-the-art methods, including up to $44.9\times$ speedup and $98.2\%$ energy savings across diverse datasets and heterogeneous hardware configurations, demonstrating the practical impact for scalable GNN deployment on wireless edge networks.

Abstract

The key to device-edge co-inference paradigm is to partition models into computation-friendly and computation-intensive parts across the device and the edge, respectively. However, for Graph Neural Networks (GNNs), we find that simply partitioning without altering their structures can hardly achieve the full potential of the co-inference paradigm due to various computational-communication overheads of GNN operations over heterogeneous devices. We present GCoDE, the first automatic framework for GNN that innovatively Co-designs the architecture search and the mapping of each operation on Device-Edge hierarchies. GCoDE abstracts the device communication process into an explicit operation and fuses the search of architecture and the operations mapping in a unified space for joint-optimization. Also, the performance-awareness approach, utilized in the constraint-based search process of GCoDE, enables effective evaluation of architecture efficiency in diverse heterogeneous systems. We implement the co-inference engine and runtime dispatcher in GCoDE to enhance the deployment efficiency. Experimental results show that GCoDE can achieve up to $44.9\times$ speedup and $98.2\%$ energy reduction compared to existing approaches across various applications and system configurations.

Figures (11)

• Figure 1: Inference speed vs. energy consumption comparison.
• Figure 2: Changes in the required transfer data size and percentage of total latency for each operation in DGCNN.
• Figure 3: Execution time breakdown of DGCNN across various devices on ModelNet40 and MR datasets.
• Figure 4: Performance of various partitioning schemes on DGCNN under different heterogeneities. Jetson TX2 serves as the device.
• Figure 5: Overview of GCoDE framework.