Task offloading for directed acyclic graph applications based on edge computing in Industrial Internet

Abstract

With an increase in the number of devices involved in the Industrial Internet, effectively combining the characteristics of industrial scenarios with an edge computing methodology for computation-intensive applications poses a critical challenge. This paper proposes an integrated architecture that allows industrial devices to offload tasks to cloud or edge servers. An offloading problem is also formulated into an energy-cost (EC) minimization problem while satisfying the deadline constraint. To solve the optimization problem, two types of offloading algorithms, namely ASO and Pro-ITGO, are proposed based on the integrated architecture. The ASO algorithm is a lightweight linear programming algorithm that includes subdeadline allocation, topology sorting, and task offloading sub-algorithms. The Pro-ITGO algorithm is a group intelligence heuristic algorithm that is derived from the original ITGO algorithm adapting the offloading scenarios of the Industrial Internet. Experimental results demonstrate that compared with state-of-the-art heuristic algorithms, the proposed algorithms can effectively reduce the energy consumption of industrial devices and cloud computing costs.

Introduction

With the promotion of intelligent manufacturing, industrial applications are facing numerous challenges such as energy consumption, quality of service (QoS) guarantee, and security, due to the inherent defects of industrial devices such as low CPU speed, constrained storage space, and insufficient sensing capability [1]. As a solution to these problems, the Industrial Internet adopts cloud computing services [2] to offload several computing tasks to cloud servers. It then uses the powerful computing capabilities of the cloud servers to accomplish tasks, thereby improving the performance of the industrial applications while ameliorating production efficiency and reducing production costs [3]. However, owing to the increase in the number of access devices, task offloading is becoming increasingly complex. The latency requirements of tasks may not be satisfied because of the large number of tasks being offloaded to cloud servers. Edge computing can improve the processing of delay-sensitive tasks by offloading these tasks to edge servers located close to the devices according to the task requirements [4].

Many industrial applications such as digital twins and automated guided vehicle (AGV) task scheduling contain a set of tasks with different priorities. Task offloading is frequently complicated because of the numerous factors that must be considered such as the deadline of the industrial applications, dependency relationships among the tasks, energy consumption of the devices, cloud computing costs involved when offloading tasks to cloud servers, and data transmission bandwidth. Firstly, owing to several factors (e.g., the constrained storage and computing capacities of the devices, and excessive time for data transmission to the cloud servers), the deadlines of industrial applications may not be met, especially for emergency tasks. Secondly, offloading tasks to cloud servers incurs a certain economic cost. Thirdly, energy consumption has an important influence on an offloading strategy, especially for energy-sensitive industrial devices such as AGVs and intelligent industrial robots. Considering the aforementioned factors in an offloading strategy comprehensively poses a critical problem. Unfortunately, previous studies have not effectively combined the characteristics of the industrial scenarios with an edge computing methodology. This study makes an initial effort toward this direction.

This subsection reviews the previous task scheduling studies on cloud and edge computing. To accelerate the execution of applications or to save storage space, an application is frequently divided into several parts called tasks that are dispatched to different computing nodes for execution. The task scheduling optimization goals in cloud computing are mainly divided into two categories, namely to minimize costs [5], [6], [7], [8], [9] and to minimize energy consumption [10], [11], [12], [13], [14]. Given that the task-scheduling problem is an NP-hard problem, the majority of the above-mentioned studies used heuristic algorithms to solve this problem. However, heuristic algorithms require multiple iterations to achieve an optimal solution, and the original algorithms must be improved to obtain stable results.

Discussions on the topic of task offloading in edge computing scenarios have recently attracted growing attention. The objectives of task offloading in edge computing scenarios differs from those in cloud computing scenarios given several metrics including delay, bandwidth resources, and energy consumption. The literature [15] proposed a dynamic programming method for the problem where the computing resources of the edge servers are difficult to manage effectively. In [16], an integer linear programming formula and two polynomial time heuristics were applied to minimize the energy consumption of smartphones. The work in [17] proposed a hybrid heuristic algorithm for task offloading among heterogeneous fog nodes of an intelligent production line. Recently, reinforcement learning technology has been applied to offloading strategies for dynamic networks. In [18], an adaptive efficient resource management algorithm was proposed to centralize the cloud and edge servers based on reinforcement learning. The literature [19] studied the offloading problem in SAGIN and proposed a learning-based method to obtain the optimal offloading strategy from the dynamic SAGIN environment. Aliyu et al. [20] proposed a moving edge computing framework based on a software-defined network and an adaptive resource-capacity management method to achieve the specified low-latency requirements of the tasks. The literature [21] adopted a fog-to-fog communication approach to reduce the service latency of sharing loads in multiple regions. However, considering the characteristics of the industrial scenarios [22] in offloading strategies introduces specific challenges.

In this paper, we focus on the task offloading problem after introducing an edge computing methodology for Industrial Internet applications in the presence of edge servers. Our main contributions are summarized as follows:

• We mainly consider the Industrial Internet scenario and focus on industrial devices that are sensitive to energy consumption. The industrial applications deployed on these devices typically have high requirements for delay and energy consumption. Therefore, a three-layer architecture of industrial devices, edge computing, and cloud computing is proposed to fully explore the capabilities of the edge computing resources in the Industrial Internet.

• We formulate the task offloading process for industrial applications deployed on the proposed three-layer architecture as a multi-objective optimization problem, and we consider minimizing the energy consumption of the industrial devices and cloud computing costs while considering the delay constraints and dependency relationships among the tasks, especially the differences in bandwidth among the layers.

• Given that a feasible set of the problem is non-convex, we design a lightweight linear programming algorithm called ASO and a group intelligence heuristic algorithm called Pro-ITGO to address the above-mentioned problem.

• We evaluate the proposed scheme using the Workflow-Sim framework and WorkflowGenerator emulator. Experimental results indicate that the proposed ASO and Pro-ITGO algorithms are capable of reducing the average energy consumption by 35% compared with state-of-the-art schemes.

The remainder of this paper is structured as follows. Section 2 addresses the system model and formulates our optimization problem. Section 3 presents the ASO and Pro-ITGO algorithms to solve this problem. Section 4 conducts extensive experiments to verify the effectiveness and efficiency of the proposed approach. Section 5 concludes the paper and presents directions for future work.