Smart computational offloading for mobile edge computing in next-generation Internet of Things networks

Abstract

Limited battery and computing resources of mobile devices (MDs) induce performance limitations in mobile edge computing (MEC) networks. Computational offloading has the capability to provide computing and storage resources to MDs for resource-intensive tasks execution. Therefore, to minimize energy consumption and service delay, MDs offload the resource-intensive tasks to nearby mobile edge server (MES) for execution . However, due to time varying network conditions and limited computing resources at MES also, the offloading decision taken by MDs may not achieve the lowest cost. In this paper, we propose an energy efficient and faster deep learning based offloading technique (EFDOT) to minimize the overall cost of MDs. We formulate a cost function which considers the energy consumption and service delay of MDs, radio resources, energy consumption and delay due to task partitioning, and computing resources of the MDs and MESs. Due to high computational overhead of this comprehensive cost function, we generate a training dataset to train a deep neural network (DNN) in order to make the decision making process faster. The proposed work finds the optimal number of components, task partitioning, and fine-grained offloading policy simultaneously. We formulate the fine-grained offloading decision problem in MEC as multi-label classification problem and propose EFDOT to minimize the computation and offloading overhead. The simulation results show high accuracy of the DNN and high performance of EFDOT in terms of energy consumption, service delay, and battery life.

Introduction

The computing requirements for emerging applications, such as wearable cognitive assistance and augmented reality applications, are increasing immensely. Such type of applications need significant battery life, memory size, computational resources, and timely execution of tasks [1]. The mobile cloud computing (MCC) [2] offers these resources, however, long distance between MD and cloud center imposes huge latency which is not acceptable for delay sensitive applications. A new distributed architecture, where small servers are placed at the edge of cellular network, is proposed by European Telecommunication Standard Institute (ETSI), known as Mobile Edge Computing (MEC) [3]. In the MEC architecture, distributed mobile edge servers (MESs) are deployed at the edge of wireless network to provide storage and computing resources with high bandwidth and low delay to enable next-generation Internet of Things networks and services. By the use of computation offloading between MDs and MESs, the communication overhead and execution delay can be reduced.

In the uncertain and time varying MEC environments, usually the MDs cannot make the most appropriate fine-grained offloading decisions in real time. Practically, not only the wireless and backhaul links between MDs and MESs are time varying and uncertain but also the resources offered by MESs are limited especially in hotspot areas [4]. A new research area, called intelligent edge learning [5], [6], has emerged recently, in which machine learning techniques are used at the network edge. The main motivation of including machine learning in MEC is to access the huge real time data generated by MDs and respond to real time offloading requirements with minimum energy consumption and service delay.

In fine-grained computational offloading, an MD divides a single task into some number of components. Now, the question arises as to, what is the optimal number of components that a task can be divided in? A task of size $M$ can be partitioned into $n$ number of components in ${}^M{Z}_n$ possible ways. For example, a 400 MB task can be divided into 3 components as: [100,100,200] MB or [200,200,0] MB or [100,300,0] MB or [400,0,0] MB. Therefore, there are four possible ways, i.e., ${}^{400}{Z}_3=4$ for 400 MB task to be divided into three components. The number of possible ways of task partitioning also depends on the minimum allowable size of a component, called division resolution, and represented as $\gamma$. For example, if division resolution is 50 MB for the same task with the same number of components, then it has 10 possible partitioning options (${}^M{Z}_n=10$), as shown in Table 1. The value of ${}^M{Z}_n$ increases exponentially with task size.

Now, the next question is, how can we find the optimal partitioning option in all ${}^M{Z}_n$ possible options? Similarly, a task of $n$ number of components can be offloaded with $2^n$ possible offloading policies. For example, a task of three components can be offloaded with $2^3=8$ possible offloading policies, as shown in Table 2. Now, the next question is, what is the optimal offloading policy in all $2^n$ possible offloading policies? To answer these three important questions, in this paper, we propose a deep learning based computational offloading technique which simultaneously finds the optimal number of components, optimal partitioning option, and optimal offloading policy, based on a comprehensive cost function. Our cost function considers the costs due to partitioning, transmission, execution, and reception to make the offloading decisions faster and, consequently, prolong the battery life of MDs. Therefore, we name our proposed technique as Energy-efficient and Faster Deep-learning based Computational Offloading Technique (EFDOT). The proposed cost function also considers the varying communication and computation resources of MES.

The proposed work considers a single MD in MEC environment to make rapid fine-grained offloading decisions with minimum cost. The novelty and contributions of the proposed work are summarized as follows:

• We consider the heterogeneous nature of task components in partial offloading which is more practical as compared to fixed number of components per task. As different types of tasks have different requirements, therefore, the number of task components vary with different types of tasks. To the best of our knowledge, this paper is the pioneering work to consider and find the optimal number of task components for different tasks.

• In partial offloading, the partitioning process is of great importance which is ignored in most of the literature on computational offloading in MEC. To the best of our knowledge, we propose the partitioning process, for the first time, for fine-grained computational offloading in MEC. The proposed work considers the cost of partitioning a task into multiple components and selects the possible partitioning option with minimum cost in all possible partitioning options.

• We combine three problems of selecting the best $\left(a\right)$ number of task components from $n$ possible options, $\left(b\right)$ task partitioning from ${}^M{Z}_n$ possible options, and $\left(c\right)$ partial offloading policy from $2^n$ possible options as multi-label classification problem. The computational overhead of finding minimum cost while considering all the three problems simultaneously becomes $O\left(n^2{}^M{Z}_n{2}^n\right)$. Therefore, to avoid this huge computation complexity, we propose EFDOT which uses deep neural network to solve all the three problems simultaneously with minimum cost value.

• We formulate a comprehensive cost function which considers multiple parameters, namely, varying network conditions and computing resources of MESs, propagation delay, the time delays and energy consumptions due to partitioning, transmission, execution, and reception.

• The proposed work minimizes the service delay and energy consumption to prolong the battery life of MDs by incorporating the applications of deep neural network (DNN) in MEC networks.

The rest of the paper is structured as follows. Section 2 reviews the related work on computational offloading in MEC networks. Section 3 formulates the mathematical model for the proposed work. Section 4 describes the proposed and benchmark techniques for partial offloading in MEC networks. Section 5 presents simulation results and Section 6 concludes the paper.