TaCo: Tasks Co-Programming for Accelerating Inference in Scaling Edge Computing

The edge computing network (ECN) accelerates edge inference by leveraging nearby resources. The method of coprogramming inference tasks on edge meets the requirement of overloaded inference on resource-limited edge devices. However, making scheduling decisions in ECN and deep neural network (DNN) two different non-Euclidean spaces is a challenging problem owing to extra scheduling complexity in dynamic scaling ECN. Here we propose a learning-based framework namely tasks co-programme (TaCo), to cooperatively accelerate the edge inference by leveraging cooperative synergy among edge devices to achieve latency-efficient services. Firstly, the high-complexity network resource arrangement (NRA) problem is transformed as a tractable graph node classification (GNC) problem according to graph optimization theory. After that, a graph convolutional network (GCN) is elaborated to jointly aggregate the network condition as well as inference workload to further enacts an intelligent arrangement scheme. Guided by a reinforcement learning (RL)-based advantage actor critic (A2C) mechanism to improve the performance of the GCN. Our results show that TaCo has 2 times faster speed to converge compared to other learning-based solutions. The decrease of inference latency is 3 times smaller than learning-based solutions, approximately 10 times smaller than traditional solutions, up to 6 times smaller than the single-UE case, and significantly lower than the singleES case, regardless of the number of nodes in the network. TaCo achieves a better energy efficiency compared to other solutions.