An Effective Offloading Trade-Off to Economize Energy and Performance in Edge Computing

Abstract

Mobile Edge Computing is a new technology which aims to reduce latency, to ensure highly efficient network operation and to offer an improved user experience. Considering offloading will introduce additional wireless transmission overhead, the key technical challenge of mobile edge computing is tradeoff between computation cost and wireless transmission cost, reducing energy consumption of mobile edge devices and response time of computation task at the same time. A Mobile Edge Computing System composed of Mobile Edge Device and Edge Cloud, connecting with Wireless Stations, comes out. To protect user privacy, Data Preprocessing is proposed which includes irrelevant property clean and data segmentation. Aimed at reducing total energy consumption and response time, an energy consumption priority offloading (ECPO) algorithm and a response time priority offloading (RTPO) algorithm are put forward, based on Energy Consumption Model and Response Time Model. Combining both ECPO and RTPO, a dynamic computing offloading algorithm is raised which is more universal. Finally, simulations in four scenarios, including network normal scenario, network congested scenario, device low battery scenario and task time limited scenario, demonstrate that our algorithms can effectively reduce energy consumption of mobile edge device and response time of computation task.

1 Introduction

With the rapid development of Internet of Things [1], mobile edge devices generate a lot of data, performing as not only data consumers but also data producers [2]. Under this circumstance, traditional cloud computing is no longer a wise choice since it has some obvious drawbacks. Firstly, it is not acceptable for real time computation task on mobile edge devices because bandwidth of network would be the bottleneck. Secondly, user privacy is another problem to be solved for cloud computing.

Mobile Edge Computing is a new technology which provides an IT service environment and cloud computing capabilities at the edge of mobile network, within the Radio Access Network and in close proximity to mobile subscribers. The aim is to reduce latency, to ensure highly efficient network operation and service delivery, and to offer an improved user experience [3]. In mobile edge computing scenario, network latency can be reduced by enabling computation and storage capacity at the edge network. And mobile edge devices can perform computation offloading for computing intensive applications to leverage the context-aware mobile edge computing service by using real time radio access network information [4]. Since offloading introduces additional communication overhead, a key technical challenge is how to balance between computation cost and communication cost to support applications with enhanced user experience, such as lower response time and energy consumption [5].

In this paper, we design a Mobile Edge Computing System composed of Mobile Edge Device and Edge Cloud, connecting with Wireless Station. To protect user privacy, Data Preprocessing is proposed including irrelevant property clean and data segmentation. Assuming that horizontal segmentation of input data will not affect computing result and execution time of computing is proportional to data size, input data can be divided into multiple data blocks and data block can be divided into multiple data slices with fixed data size. Only one data slice can be executed each time. In our research, we focus on reducing energy consumption of mobile edge device and response time of computation task, proposing an Energy Consumption Model and a Response Time Model.

In terms of mobile edge computing offloading problem, considering the change of network environment and the power of mobile edge device to make a dynamic decision will work better than making an unchanging decision at the beginning. Therefore, an energy consumption priority offloading (ECPO) algorithm and a response time priority offloading (RTPO) algorithm are raised in order to reduce energy consumption and response time. Furthermore, combining both ECPO algorithm and RTPO algorithm, we propose a new dynamic computing offloading algorithm as the greatest contribution of our paper. Finally, simulations in four scenarios: network normal scenario, network congested scenario, device low battery scenario and task time limited scenario, demonstrate that our proposed algorithms can effectively reduce energy consumption and response time.

The structure of this paper is as follows. Section 2 presents the related work. Section 3 describes system design and system model. In Sect. 4, we propose ECPO alogorithm, RTPO algorithm and a dynamic computing offloading algorithm combining previous two. Section 5 shows the simulation results and analysis. Section 6 is about conclusion and future work.

2 Related Work

Recently, computing offloading from mobile devices into cloud [6], as a key technical challenge of Mobile Edge Computing [7], has been extensively studied in many area including power management [8], cloud computing task migration [9, 10] and virtual machine migration [11].

In order for better understanding of offloading, Orsini et al. provide a design guideline for the selection of suitable concepts for different classes of common cloud-augmented mobile applications and present open issues that developers and researchers should be aware of when designing their mobile cloud computing approach [12].

There are some existing offloading algorithms, like energy-optimal partial computation offloading (EPCO) algorithm [5], Lyapunov optimization-based dynamic computation offloading (LODCO) algorithm [13], distributed computation offloading algorithm [14] and the actor-model programming paradigm [15]. Sardellitti et al. consider an MIMO multicell system where multiple mobile users ask for computation offloading to a common cloud server and propose an iterative algorithm, based on a novel successive convex approximation technique, converging to a local optimal solution of original nonconvex problem [16].

In order to save energy, Zhao et al. design a threshold-based policy to improve the QoS of Mobile Cloud Computing by cooperation of local cloud and Internet cloud resources, which takes advantages of low latency of local cloud and abundant computational resources of Internet cloud simultaneously [17]. Ge et al. propose a game-theoretic approach to optimize the overall energy in a mobile cloud computing system and formulate the energy minimization problem as a congestion game, where each mobile device is a player to select one server to offload computation to minimize the overall energy consumption [18]. Wang and Giannakis investigate resource allocation policies for time-division multiple access over fading channels in the power-limited regime [19].

What’s more, Chen et al. study the multi-user computation offloading problem for mobile-edge cloud computing in a multi-channel wireless interference environment and design a distributed computation offloading algorithm that can achieve a Nash equilibrium, deriving the upper bound of the convergence time and quantifying its efficiency ratio over the centralized optimal solution in terms of two important performance metrics [14]. You et al. study resource allocation for a multiuser MECO system based on time-division multiple access (TDMA) and orthogonal frequency-division multiple access (OFDMA) to solve the problem and characterize its policy structure, proposing a low-complexity sub-optimal algorithm by transforming the OFDMA problem to its TDMA counterpart [20].

3 System Design and System Model

In this section, we introduce system architecture design and data preprocessing method. The energy consumption model and response time model are also presented. Partial parameters used in the following are listed in Table 1.

Table 1. Partial parameters

3.1 System Architecture Design

The system is composed of two requisite parts: Mobile Edge Device and Edge Cloud, while data transmission of them is Wireless Station. The system architecture is shown in Fig. 1.

Fig. 1.

System architecture

Mobile Edge Device. Mobile Edge Device is both data consumer and data producer in edge computing, including three modules: Data Preprocess Module, Computing Module and Offloading Module. Data Preprocess Module is responsible for irrelevant property clean and data segmentation. Computing Module is used for local computing. Offloading Module is used to develop calculation migration strategy.

Edge Cloud. There are two modules in Edge Cloud: Computing Execution Module and Data Collection Module. Computing Execution Module is a module that performs computational tasks, while Data Collection Module is applied for collecting data from Mobile Edge Device.

3.2 Data Preprocessing

There are two kinds of data preprocessing: Irrelevant Property Clean and Data Segmentation. On the one hand, the original data may contain some properties which are related to user private and computing irrelevant. On the other hand, in the case of horizontal segmentation of input data not affecting computing result, we propose to segment from multiple granularities on the basic definition of Data Unit, smallest data processing unit for performing computational tasks. Data Unit can be divided into multiple Data Blocks while Data Block can be divided into multiple Data Slices as shown in Fig. 2. The purpose of multi-granularity data segmentation is to dynamically adjust the computing migration mechanism for different granularity data which will be analyzed in Sect. 4.

Fig. 2.

Data segmentation

3.3 Energy Consumption Model

One of the main purposes of mobile edge computing is to reduce energy consumption of mobile edge devices as much as possible, including energy consumption for local computing and for wireless transmission.

Energy Consumption for Local Computing. For a mobile edge device, let $E_{c}$ denote the energy consumption per CPU cycle for local computing, $C_{m}$ denote the number of CPU cycles required for computing 1-bit of input data and $D_{l}$ denote the size of data slice. Then the total energy consumption for local computing of a data slice is $E_{l}$, given by Formula 1.

$$\begin{aligned} E_{l}=C_{m}\cdot D_{l}\cdot E_{c} \end{aligned}$$

(1)

Energy Consumption for Wireless Transmission. Let W denote the channel bandwidth, g denote the channel gain, $P_{t}$ denote the transmission power and N denote the variance of complex white Gaussian channel noise. Then the channel transmission rate denoted by $R_{t}$, is Formula 2, according to Shannon formula.

$$\begin{aligned} R_{t}=W\cdot \log _{2}(1+\frac{g\cdot P_{t}}{W\cdot N}) \end{aligned}$$

(2)

Provided that W, $P_{t}$ and N are constant, $R_{t}$ will have a positive correlation with changeable g. However, if time slot is short enough, $P_{t}$ and $R_{t}$ can be considered constant, resulting to the total energy consumption for wireless transmission of a data slice as $E_{t}$, given by Formula 3, where $T_{t}$ and $D_{t}$ are transmission time and size of offloading data slice respectively.

$$\begin{aligned} E_{t}=P_{t}\cdot T_{t}=\frac{P_{t}\cdot D_{t}}{R_{t}} \end{aligned}$$

(3)

By the way, energy consumption for wireless transmission of calculation result is negligible in this case due to size of result is much smaller than that of original data normally.

Total Energy Consumption. Considering the size of data slice is small enough, the local computing time and transmission time of data slice is short. Under this circumstances, analysis of energy consumption is acceptable.

Let D denote size of the whole data slice, $D_{l}$ denote the size of data slice in local computing with $0\le D_{l}\le D$, then the size of offloading data slice is $D_{t}=D-D_{l}$. Total energy consumption denoted as E is given by Formula 4.

$$\begin{aligned} E=E_{l}+E_{t}=\frac{P_{t}}{R_{t}}\cdot D+(C_{m}\cdot E_{c}-\frac{P_{t}}{R_{t}})\cdot D_{l} \end{aligned}$$

(4)

For the case of $C_{m}\cdot E_{c}-\frac{P_{t}}{R_{t}}>0$, E will be minimum when $D_{l}=0$. Oppositely, supposing $C_{m}\cdot E_{c}-\frac{P_{t}}{R_{t}}\le 0$, E will be minimum when $D_{l}=D$. It shows that for a data slice small enough, to make energy consumption minimum, the whole data slice is either executed in local computing or offloaded to edge cloud.

The total energy consumption of one computation task can be expressed as Formula 5.

$$\begin{aligned} E_{total}=\sum _{}^{}E_{slice}, \ for \ each \ data \ slice \end{aligned}$$

(5)

3.4 Response Time Model

Another purpose of mobile edge computing is to reduce computing response time, which is mainly affected by three factors: local computing, wireless transmission and cloud computing.

Response Time for Local Computing. For a mobile edge device, let $f_{m}$ denote CPU frequency for local computing which means the number of CPU cycles per second, $C_{m}$ denote the number of CPU cycles required for computing 1-bit data and $D_{l}$ denote the size of data slice in local computing. Then the total response time for local computing of a data slice is $T_{l}$, given by the following Formula 6.

$$\begin{aligned} T_{l}=\frac{C_{m}\cdot D_{l}}{f_{m}} \end{aligned}$$

(6)

Response Time for Wireless Transmission. If the time slot is short enough, the channel transmission rate $R_{t}$ can be considered to be constant. In this case, the total response time for wireless transmission of a data slice of input data is $T_{t}$, given by Formula 7.

$$\begin{aligned} T_{t}=\frac{D_{t}}{R_{t}}=\frac{D-D_{l}}{R_{t}} \end{aligned}$$

(7)

where $D_{t}$ is the size of offloading data slice. Response time for wireless transmission of computing results is ignored due to the relative smaller sizes.

Response Time for Cloud Computing. Obviously in edge cloud, the CPU frequency denoted as $f_{e}$ and the number of CPU cycles required for computing 1-bit data denoted as $C_{e}$ can be considered constant within a time slot which is short enough. Then the total response time for cloud computing of a data slice is $T_{e}$, given by Formula 8.

$$\begin{aligned} T_{e}=\frac{C_{e}\cdot D_{t}}{f_{e}}=\frac{C_{e}\cdot (D-D_{l})}{f_{e}} \end{aligned}$$

(8)

Total Response Time. For a data slice, local execution and offloading to edge cloud are simultaneous, which leads to the total response time be the maximum of them, given by Formula 9.

$$\begin{aligned} T=max(T_{l},T_{t}+T_{e}) \end{aligned}$$

(9)

Total response time of a computation task can be expressed as Formula 10.

$$\begin{aligned} T_{total}=\sum _{}^{}T_{slice},\ for\ each\ data\ slice \end{aligned}$$

(10)

4 Computing Offloading Mechanism

Reducing energy consumption and response time is two main purpose for computing offloading mechanism in edge computing. In this section, we proposed an energy consumption priority offloading algorithm named ECPO and a response time priority offloading algorithm named RTPO. Ultimately, we put forword a dynamic computing offloading algorithm based on ECPO and RTPO.

4.1 Energy Consumption Priority Offloading Algorithm

To reduce the total energy consumption of mobile edge device, an energy consumption priority offloading (ECPO) algorithm is proposed in Algorithm 1.

As is shown in Fig. 3, in ECPO algorithm, data will be divided into multiple data slices with fixed data size under the assumption that horizontal segmentation will not affect computing result. For each data slice, it will be either executed in local computing or offloaded to edge cloud. Only one data slice can be executed each time. What’s more, total energy consumption of one data slice E in ECPO algorithm is shown in Formula 4. As is proved by Energy Consumption Model, E will be minimum when using ECPO algorithm.

Fig. 3.

Energy consumption priority offloading

4.2 Response Time Priority Offloading Algorithm

Considering that ECPO algorithm has space for improving response time of computation task, an response time priority offloading (RTPO) algorithm is proposed in Algorithm 2.

As is shown in Fig. 4, in RTPO algorithm, the offloading weight of a data slice is calculated according to the power percentage of mobile edge device and network status. In other words, RTPO algorithm allows one data slice to execute both in local and edge cloud to earn less response time.

Fig. 4.

Response time priority offloading

For each data slice, the response time for executing local computing to process a data unit is $\frac{C_{m}}{f_{m}}$, while time for offloading to edge cloud is $\frac{1}{R_{t}}+\frac{C_{e}}{f_{e}}$. So for a data slice, in order to equalize this two latencies, we get the weight for local computing $D_{l}$ in Formula 11. As is analyzed previously in Formula 4, total energy consumption of one data slice E in RTPO algorithm will not be the worst case as $D_{l}$ always satisfy the condition $0<D_{l}<D$.

$$\begin{aligned} D_{l}=\frac{(f_{m} \cdot f_{e} + C_{e} \cdot f_{m} \cdot R_{t}) \cdot D}{C_{m} \cdot f_{e} \cdot R_{t} + f_{m} \cdot f_{e} + C_{e} \cdot f_{m} \cdot R_{t}} \end{aligned}$$

(11)

4.3 Dynamic Computing Offloading Algorithm

In order to deal with more complex scenarios, we propose a dynamic computing offloading algorithm in Algorithm 3, combining ECPO and RTPO algorithm.

The dynamic computing offloading algorithm takes both power of mobile edge device and user requirements into consideration. On the one hand, if mobile edge device does not have enough power to support, task will be abandoned. Besides, let $threshold\_power$ denote the threshold of percentage of battery charge and compare power before processing each data block. On the other hand, let $w_{e}$ and $w_{t}$ denote the weight of reducing energy consumption and shortening response time respectively according to user requirements to calculate overhead as Formula 12. Finally, we select algorithm with lower overhead to deal with current data block.

$$\begin{aligned} Overhead = E \cdot w_{e} + T \cdot w_{t} \end{aligned}$$

(12)

The energy consumption model and time model of one computaion task are the sum of cost in each data block, which can be expressed as Formula 13.

$$\begin{aligned} \begin{aligned}&E_{total}=\sum _{}^{}E_{block},\ for\ each\ data\ block\\&T_{total}=\sum _{}^{}T_{block},\ for\ each\ data\ block \end{aligned} \end{aligned}$$

(13)

5 Evaluation

As 5G technology has not fully applied currently, it is scarcely possible to carry out experiments in real environment. In this section, we analyze our simulation result in four scenarios to evaluate performance of the proposed dynamic computing offloading mechanism, including network normal scenario, network congested scenario, device low battery scenario and task time limited scenario.

5.1 Experimental Parameter

Settings of our simulation are as follows. The size of data block and data slice are 100 MB and 1 MB separately. Energy consumption per CPU cycle for mobile edge device $E_{c}=1.0\times 10^{-10}$ J/cycle. The number of CPU cycles required for computing 1-bit data at mobile edge device and edge cloud is $C_{m}=C_{e}=1000$ cycles/bits. CPU frequency for a mobile edge device is $f_{m}=2.0\times 10^{9}$ cycles/s while it for edge cloud is $f_{e}=1.0\times 10^{10}$ cycles/s. Transmission power is $P_{t}\in [0, 0.2]$ J/s. Channel transmission rate is $R_{t}\in [0,2000]$ KB/s. Energy of mobile edge device at full charge is 20000 J and $threshold\_power$ is set to $30\%$.

Moreover, as Formula 14 shows, WirelessStation consists of n channels and m devices are maintained for each channel, of which m has negative correlation of the wireless transmission rate.

$$\begin{aligned} \begin{aligned}&WirelessStation(Channel_{1},Channel_{2},...,Channel_{n-1},Channel_{n})\\&Channel(Device_{1},Device_{2},...,Device_{m-1},Device_{m}) \end{aligned} \end{aligned}$$

(14)

Assuming that $Number_{Channel}$ and $Number_{Device}$ represent the number of channels and mobile edge devices severally, we define the ratio of $Number_{Channel}$ and $Number_{Device}$ equalizing 20 as congested network, while equalizing 2 as normal network in our experiment.

5.2 Simulation Results Analysis

Network Normal Scenario. Processing computation task with size of 100 MB in normal network scenario with $Number_{Device}/Number_{Channel}=2$, total energy consumption and response time of edge device are shown in Fig. 5. Apparently, ECPO, RTPO and Dynamic Algorithm have a great advantage over executing locally and all offloading to edge cloud. However, in terms of energy consumption, compared with ECPO and RTPO, advantage of Dynamic Algorithm is not obvious, as even slightly worse than ECPO. While in terms of response time, combining the superiority of RTPO, it performs better for Dynamic Algorithm than ECPO while a little worse than RTPO. Therefore, Dynamic Algorithm plays a role as a compromise between ECPO and RTPO.

Fig. 5.

Analysis in normal network scenario

Network Congested Scenario. Dealing with task with size of 100 MB in congested network scenario with $Number_{Device}/Number_{Channel}=20$, results are shown in Fig. 6. In the respect of energy consumption, cost of Dynamic Algorithm is slightly higher than local execution due to high transmission cost in congested network, while is half as low as it of all offloading. From another perspective, response time of Dynamic Algorithm has distinct improvement over no offloading and all offloading, but still in the midst of ECPO and RTPO.

Fig. 6.

Analysis in Congested Network Scenario

Device Low Battery Scenario. The initial power of mobile edge device is set lower than $30\%$ randomly in low battery scenario and the original data size of each task is 5 GB. Number of initial tasks is 100. Results of computing offloading is shown in Table 2. The failure of computing tasks is due to exhaustion of mobile edge devices. Success rates of ECPO and RTPO are $91\%$ and $75\%$ each. Dynamic Algorithm will estimate energy consumption before executing tasks and abandon calculation task when do not have enough power, resulting to $100\%$ success rate.

Table 2. Result of computing offloading in device low battery scenario

Task Time Limited Scenario The initial power of mobile edge device is set randomly and the original data size of each task is 5 GB. Number of initial tasks is 100. In addition, each computing task is limited to 1 h. Results of computing offloading is shown in Table 3. For ECPO, there are 2 tasks failed due to depletion of mobile edge devices, while the remaining $98\%$ violate the time limit due to timeout. For RTPO, although its success rate is $95\%$, 5 tasks failed due to exhaustion of mobile edge device. For Dynamic Algorithm, success rate is $83\%$ without failure which is optimal entirely.

Table 3. Result of computing offloading in task time limited scenario

5.3 Brief Summary

The simulation results conducted in 4 scenarios show that our computing offloading machanism can comprehensively consider energy consumption, response time of computing tasks and load of edge cloud, by dynamically adjusting and calculating offloading strategy. The machanism is far better than local execution and all offloading to edge cloud.

6 Conclusion

We study computing offloading mechanism for a mobile edge computing system composed of mobile edge device and edge cloud, connecting with wireless station, to reduce energy consumption of mobile edge devices and response time of computation tasks. We propose ECPO and RTPO algorithm according to Energy Consumption Model and Response Time Model. Ultimately, we put forward a dynamic offloading algorithm combining the previous two, which is proved to be an effective way to achieve the original goals entirely. In the future, after full application of 5G technology, we will make further effort to build a real mobile edge computing scenarios to verify the effectiveness of our mechanism.