With the development of communication and networking, the number of mobile users is increasing in a staggering speed. According to Oracle, there are more than 10 billion mobile devices connected to the Internet, and that number will grow to 22 billion by 2025 [1]. The explosive growth of mobile devices leads to a large request and the promising application prospect [2]. New technologies are beginning to be applied to mobile devices, such as virtual reality (VR), artificial intelligence (AI) and the Internet of Vehicles (IoV) [3]. Among these, as a prospective technology, VR is constantly enriching people's experience ( [4], [5]). In practice, the VR applications are usually run over a wearable VR devices (VDs), which can record information about the movement and activity of VR users (VUs). Equipment manufacturers of VD usually need to consider physical size constraint, which equips with a small CPU and GPU with low-power consumption and low-capacity battery [6]. However, due to VR application requires real-time processing, it will consume a lot of computing resources. This poses a big challenge to VDs.
Fortunately, mobile edge computing (MEC) is a CEUR Workshop Proceedings (CEUR-WS.org)
promising distributed computing paradigm ( [7], [8]). Computation offloading of MEC has been well studied in ( [9]- [10]). Inspired by this, offloading the VR application to edge servers (ESs), the time and energy consumption of VD can be significantly reduced. However, computing resources of the ES are limited, it will increase the waiting time and energy consumption of VDs when the number of tasks performed exceeds computing capacity of the ES [11]. Meanwhile, the resources of ESs are heterogeneous, the tasks should be reasonably distributed on ES cluster to avoid some of the ESs are overload [12]. Meanwhile, when the VU interacts with the VD, the application will collect the VU's private information, such as location, posture [13]. If placing data with privacy conflicts on the same ES, it is easy to cause the leakage of VU privacy information [14]. Therefore, it is necessary to consider the load balancing and the privacy constraints of placement of computation offloading for VR application.
To address above issues, the computation offloading considering privacy protection for VR applications are investigated in this work. The contributions of this paper can be summarized as follows.
1) We model the computation offloading problem as a multi-objective optimization issue, where the motionto-photons latency and energy consumption of VD, as well as load balancing of ES are considered as the optimization objectives, and privacy protection is considered as a constraint.
2) A computation offloading method for VR applications based on multi-objective optimization genetic algorithm, named MCOVR, is proposed to address the above issue. 3) We conduct a large number of comparative experiments to prove that MCOVR can obtain the optimal computation offloading strategy for VR application.
The remaining of the paper is describe as follows. Firstly, section 2 describes the system model and defines the problem formulation. Secondly, section 3 introduces our proposed multi-objective computation offloading of VR application method. Then, section 4 demonstrates the experiment results. Finally, section 5 introduces the related work and section 6 describes the conclusion and the future work.
In this section, the architecture of MEC-enabled VR application is firstly introduced. Then, the model of VR application is established. Finally, the system model and problem formulation are described.
The MEC-enabled VR application architecture is shown in Figure 1. We assume that the cloud has infinite computing resources, and resources of ESs are finite and heterogeneous. These VR applications are executed by VDs or offloaded to the ES or to the cloud. VUs communicate with ESs via local area network (LAN) and with the cloud via wide area network (WAN).
Y is a set of VU, which is defined as Y = {1, 2, ..., y}. There are vn applications in V , which is defined as V = {v 1 , v 2 , ..., v vn }. The task length requested by the v-th application of y-th user is denoted as ly,v, and the task workload is denoted as ry,v. Let X be the set of ES, denoted as X = {1, 2, ..., x}. In each ES, the computing resources of ES are represented as a number of i virtual machine(VM) xn, represented as xn = {xn 1 , xn 2 , ..., xn i }. J is a set of computation offloading strategies, which is defined as J = {j1,1, ..., j1,v, j2,1, ..., jy,v}. And jy,v represents the offloading strategies of v-th application from y-th VU. jy,v = 0 represents that the application is run by VD, jy,v ∈ (1, 2, ..., x) indicates the application is offloaded to ES, jy,v = x + 1 represents the application is offloaded to cloud for processing.
In general, Augmented reality (AR) application can be represented as a directed acyclic graph(DAG) [15]. Similarly, VR can also be represented by a DAG. As shown in Figure 2, the VR model can be represented as follows. The VR application contains three independent operations, namely, rendering, calibration and processing. Thus, we use a layer containing three operations to represent each video frame, namely node. In addition, there are k video frames in a VR application, the node k represents the k−th node in application.
Execution latency is the time cost of processing VR applications on the three platforms, namely, VD, ESs, and cloud center. In this part, the executing time of v-th VR application of y-th user is represented as
(1)
The transmission latency is related to the computation offloading strategy of the two tasks. Let P c (c ∈ {1, 2, 3}) be the flag between two strategies, which are defined as (2) According to transmission strategy P c , the transmission time S t (jy,v) of the v-th application of the y-th VU is shown as
where B represents the transmission bandwidth. The transmission bandwidth is divided into three cases which is denoted as
When the number of applications offloaded to the ES exceeds the number of VM xn, the newly coming tasks need to wait for the previous task to complete execution. The i-th VM in x-th ES is represented as a double-tuple xn x i = (vmwk, tnum), where vmwk represents the ith VM collection and the tnum represents the number of applications in VM. When the offloading strategy jy,v = x, ry,v is offloaded to VM of x-th ES. Then, the vmwk is updated by vmwk = vmwk + ly,v and the tnum = tnum + 1. Finally, the S w (jy,v) of v-th application from y-th VU is calculated as
where ψy,v,q is to determine whether ry,v needs to wait for the previous task, and the execution latency of previous task is represented as S e (pre(jy,v)).
The total time S(jy,v) is represented as the sum of all time, including the executing latency S e (jy,v), the transmission latency S t (jy,v) , and the waiting latency S w (jy,v). The total latency is represented as S(jy,v) = S e (jy,v) + S t (jy,v) + S w (jy,v). (6)
The executing energy of MDs is related to the executing time, the executing time of v-th VR application of y-th user is represented as
where ℘ active represents active state of VD energy consumption and ℘ idle represents idle state of VD energy consumption.
N t () represents transmission energy consumption, which can be obtained by multiplying the transmission power by the transmission time. N t () can be calculated by
where ℘ trans represents the energy consumption of transmission.
Waiting energy consumption is the energy consumption of VD when the task waits for the execution of the previous task. Waiting energy consumption is related to waiting time which can be obtained as
The total energy consumption N (jy,v) is represented as the sum of all energy consumptions which includes the executing energy consumption N e (jy,v), the transmission energy consumption N t (jy,v), and the waiting energy consumption N w (jy,v). The total energy consumption can be calculated by N (jy,v) = N e (jy,v) + N t (jy,v) + N w (jy,v). (10)
The ESs are heterogeneous means that the computing resources of ESs are not equal. Therefore, the workload of ESs needs to be balanced. And, ζv is a binary number to calculate the occupation of ES, which is calculated by
And, ηv is a flag to represent whether the ES is occupied, which is represented as
Firstly, the number of ES that are utilized is defined as UE(j), which is represented as
τ represent the amount of occupied ESs which is calculated by
According to the resource utilization of each ES, the average resource utilization α(j) can be obtained by
Finally, the load balancing of each ES is calculated by the squared difference between the resource utilization of each ES and the average resource utilization. The loading balancing F (J) is calculated by
As VR applications need to collect user information, VR privacy conflicts need to be considered. In this paper, we address privacy conflicts by placing applications on different ESs to avoid privacy leakage [14]
The main objectives of this study are to decrease the motion-to-photons latency and energy consumption of VDs, as well as the load balancing of ESs while preserving the VU's privacy. The problem formulations are defined as follows.
In this section, we describe the details of our proposed multi-objective computation offloading of VR application (MCOVR). MCOVR is based on MOMBI [16].
The initialization of the algorithm includes the definition of key parameters and the generation of the initial population. Firstly, the first-generation population Z0 of size t z is generated randomly. Secondly, MCOVR defines some parameters, such as crossover probability L c , mutation probability L m , the number of iterations Dmax, and current iteration index d. Finally, the algorithm defines a size of t q achieve set Q0 that keeps the achieve solution. Moreover, tournament selection, crossover, as well as mutation is executed on Z0 to generate new population Z * 0 .
In the crossover operation, the algorithm randomly exchanges a certain point value of two offloading strategies. Through crossover operation, the algorithm can obtain diversity solutions. In the mutation operation, mutations will slightly change certain values on the offloading strategy. And mutation operator can reduce the occurrence of local optimal and avoid early convergence.
In the MCOVR, algorithm divides the population through the R2 indicator to achieve non-dominated sorting, namely, R2 ranking. In R2 indicator, the objectives are measured based on the weighted Tchebycheff method and processed by normalization. W is represented as the weight of each object, A is represented the Pareto solution set, and ϑ represents the utility functions. Therefore, the R2 indicator is defined as follows.
The two populations Z d and Z * d generated by the algorithm need to select the number of t z as the nextgeneration population Z d+1 , where d<Dmax. In addition, based on R2 ranking of Z d , by comparing the objective function values of the two solutions and select the optimal value of the objective function as the parent solution.
This method aims to reduce the time consumption and energy consumption of VDs as well as load balancing of ESs jointly. The flow chart represents the execution sequence of our proposed MCOVR is shown in Figure 3.
In this section, we present the experimental evaluation of our method. Firstly, the experiment setting is introduced, which includes comparative methods and key parameters. Then, the experimental results and discussion are described.
Two comparative computation offloading methods are proposed. The details of the two methods is as follows.
Benchmark: All applications are offloaded to the three platforms for processing randomly, named as Benchmark.
First Come First Service (FCFS): All applications are offloaded to the three platforms for processing orderly, named as FCFS.
The key parameter value is described in Table 1. We execute these offloading methods by JAVA over Win10 64 OS with 8 Intel Core i7-10850H 3.60 GHz CPU processors, Nvidia RTX2070 GPU processors, and 32GB RAM.
We design three different kinds of comparative experiments to verify the effectiveness of MCOVR. All the experiments are performed 10 times, and the result is the average value.
The comparison of motion-to-photons latency is presented in Figure 4. Figure 4(a) testifies the motion-tophotons latency of multi-VU condition. As the number of VUs increases, the difference of the latency consumed among the three methods is getting larger. Our proposed
Many efforts have been make to solve the computation offloading in MEC. Bi et al. [9] considered energy consumption with the maximum number of memory constraint in MEC. An efficient algorithm based on a penalty function method is proposed to handle constraint. Xu et al. [10] jointly considered both IoV energy consumption and load balancing of ES problem. They proposed a genetic algorithms-based method to solve it. Feng et al. [17] considered the computation performance for offloading in the wireless powered MEC. Li et al. [18] studied a joint optimization problem of computation offloading and privacy preservation in the MEC system and proposed a privacy-aware online learning algorithm to solve this challenge. Ko et al. [19] studied the computation offloading problem in smart cities based on the privacy entropy model. A constrained Markov decision process method is developed to solve location privacy leakage. The above-mentioned studies provided a good view of the computation offloading in MEC and inspired us to study the computation offloading of VR applications.
A few studies take into account MEC-enabled VR applications. Zhang et al. [20] proposed an edge-cloud architecture of VR application that frame rendered on ES or cloud. They proposed a service placement algorithm for multi-user VR applications that makes placement decisions, based on QoS and VUs mobility patterns. Zhu et al. [21] focused on dynamic rendering-module placement problem for VR streaming. To reduce such rendering latency, they proposed a prediction-based pre-rendering mechanism at the ES. In [22], the authors investigated the problem of providing service using the MEC network architecture with a frame rendering. They solved an online service placement problem by leveraging model predictive control and overcoming conflicting system objectives. In [23], the authors focused on the online offloading interactive VR application. They considered the constraints of the delay requirements, the maximum frame per second demands, the total bandwidth and rendering resources limits and proposed programming to optimize execution efficiency of VR application.
Different from the above studies, we investigates how to make privacy-aware computation offloading decisions to reduce the motion-to-photons latency and energy of VR applications and load balancing of ESs while considering VUs privacy-preserving. Correspondingly, we solve this issue using a multi-objective optimization method.
We studied the problem of multi-objective computation offloading of VR applications. Correspondingly, we propose a method named MCOVR to minimize the motion-to-photons latency and energy consumption of VD and the load balancing of ESs jointly. We also considered privacy conflicts in VR applications and using the privacy placement model to protect user data information security. Finally, the experimental results have shown that our proposed method is effective.
In future work, we will study mobility-aware computation offloading for augmented reality applications in smart cities.
This work is supported by the National Science Foundation of China (Grant No.61902133), the Natural Science Foundation of Fujian Province (Grant No.2018J05106), the Fundamental Research Funds for the Central Universities(ZQN-817), Quanzhou Science and Technology Project(No.2020C050R).