ICTA: Intelligent Computing Task Allocation for Efficient Deep Learning in Distributed Edge Computing System of IoT

Abstract

As a critical part of distributed cross-device edge computing systems, computing task allocation is responsible to reasonably map the computation of DL tasks onto the set of available Internet-of-Things (IoT) devices, with the aim to achieve efficient execution of deep learning (DL) model inference/training. The complexity and uncertainty of task scheduling on diverse IoT edge devices, as well as the dynamic and resource-constrained edge computing environment, make it hard to obtain the best task allocation strategy in terms of the optimal matching between DL workloads and available resources. In order to maximize resource utilization and optimize task allocation, in this paper, we propose Intelligent Computing Task Allocation (ICTA), which is an automatic end-to-end optimizing model to allocate proper resources for each operator node in a computation graph by learning the long-term optimal resource management and task scheduling strategies. ICTA is capable of extracting features from resource graph and computation graph, respectively, by using graph convolutional network (GCN), and subsequently predicting the system performance of a given task allocation strategy through deep neural network (DNN) based on the extracted features. Finally, ICTA decides which device in the resource graph to place the operator node in the computation graph, based on the task allocation strategy corresponding to the best system performance. Moreover, being trained periodically in an end-to-end manner according to a continuous learning mechanism, GCN-based ICTA will become smarter while being used. Therefore ICTA facilitates the realization of intelligent distributed edge computing system and further contributes to smart edge applications and services.