Auto-redirect paused manually.
Latency-Aware Edge-Cloud Collaboration for Distributed AI Inference: A Convex Optimization Perspective | Multidisciplinary Research in Computing Information Systems
Abstract
Abstract
The proliferation of Internet of Things devices and the subsequent demand for intelligent applications have catalyzed the shift from centralized cloud computing to distributed edge computing paradigms. This paper addresses the critical challenge of minimizing inference latency in distributed Artificial Intelligence systems through a collaborative edge-cloud architecture. While deep neural networks provide high accuracy for tasks such as image recognition and natural language processing, their computational demands often exceed the capabilities of resource-constrained edge devices. Conversely, offloading all tasks to a centralized cloud introduces significant wide-area network transmission delays. To navigate this trade-off, we propose a mathematically rigorous framework based on convex optimization to determine optimal task offloading ratios and resource allocation strategies. We formulate the latency minimization problem as a joint optimization of computation resources, transmit power, and offloading decisions, subject to strict power and bandwidthconstraints. By relaxing the integer constraints inherent in binary offloading decisions, we transform the problem into a convex form solvable via standard optimization techniques. Our theoretical analysis and numericalsimulations demonstrate that the proposed method significantly reduces system-wide latency compared to local-only, cloud-only, and heuristic-based approaches. This work provides a scalable foundation for real-time AI inference in next-generation networks.
The proliferation of Internet of Things devices and the subsequent demand for intelligent applications have catalyzed the shift from centralized cloud computing to distributed edge computing paradigms. This paper addresses the critical challenge of minimizing inference latency in distributed Artificial Intelligence systems through a collaborative edge-cloud architecture. While deep neural networks provide high accuracy for tasks such as image recognition and natural language processing, their computational demands often exceed the capabilities of resource-constrained edge devices. Conversely, offloading all tasks to a centralized cloud introduces significant wide-area network transmission delays. To navigate this trade-off, we propose a mathematically rigorous framework based on convex optimization to determine optimal task offloading ratios and resource allocation strategies. We formulate the latency minimization problem as a joint optimization of computation resources, transmit power, and offloading decisions, subject to strict power and bandwidthconstraints. By relaxing the integer constraints inherent in binary offloading decisions, we transform the problem into a convex form solvable via standard optimization techniques. Our theoretical analysis and numericalsimulations demonstrate that the proposed method significantly reduces system-wide latency compared to local-only, cloud-only, and heuristic-based approaches. This work provides a scalable foundation for real-time AI inference in next-generation networks.