As a new technology, virtual reality (VR) is constantly enriching people's experience. VR application has high requirements for the performance VR devices, but the computing resources of VR devices are limited. Mobile edge computing provides an efective solution to solve the above issue by ofloading VR application to edge servers (ESs) for processing. Nevertheless, the resources of ESs are limited and heterogeneous. And thus, it is necessary to consider the load balancing of ESs. In addition, user privacy protection in the process of computation ofloading is another issue that needs to take into consideration. In view of this, in this paper, we investigate the computation ofloading for VR applications with privacy protection. Technically, we propose a privacy-aware computation ofloading method based on multi-objective optimization genetic algorithm to obtain the optimal strategy for VR application. Finally, it is shown that the proposed method is efective through extensive experiments.
promising distributed computing paradigm ([7],[8]).
Computation ofloading of MEC has been well studied in
With the development of communication and net- ([9]-[10]). Inspired by this, ofloading the VR application working, the number of mobile users is increasing in a to edge servers (ESs), the time and energy consumption staggering speed. According to Oracle, there are more of VD can be significantly reduced. However, computing than 10 billion mobile devices connected to the Internet, resources of the ES are limited, it will increase the waiting and that number will grow to 22 billion by 2025 [1]. The time and energy consumption of VDs when the number explosive growth of mobile devices leads to a large re- of tasks performed exceeds computing capacity of the ES quest and the promising application prospect [2]. New [11]. Meanwhile, the resources of ESs are heterogeneous, technologies are beginning to be applied to mobile de- the tasks should be reasonably distributed on ES cluster vices, such as virtual reality (VR), artificial intelligence to avoid some of the ESs are overload [12]. Meanwhile, (AI) and the Internet of Vehicles (IoV) [3]. Among these, when the VU interacts with the VD, the application will as a prospective technology, VR is constantly enriching collect the VU’s private information, such as location, people’s experience ([4],[5]). In practice, the VR applica- posture [13]. If placing data with privacy conflicts on the tions are usually run over a wearable VR devices (VDs), same ES, it is easy to cause the leakage of VU privacy inwhich can record information about the movement and formation [14]. Therefore, it is necessary to consider the activity of VR users (VUs). Equipment manufacturers load balancing and the privacy constraints of placement of VD usually need to consider physical size constraint, of computation ofloading for VR application. which equips with a small CPU and GPU with low-power To address above issues, the computation ofloading consumption and low-capacity battery [6]. However, due considering privacy protection for VR applications are to VR application requires real-time processing, it will investigated in this work. The contributions of this paper consume a lot of computing resources. This poses a big can be summarized as follows. challenge to VDs. 1) We model the computation ofloading problem as Fortunately, mobile edge computing (MEC) is a 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 ofloading 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 ex- the ofloading strategies of v-th application from y-th periments to prove that MCOVR can obtain the optimal VU. jy,v = 0 represents that the application is run computation ofloading strategy for VR application. by VD, jy,v ∈ (1, 2, ..., x) indicates the application is
The remaining of the paper is describe as follows. ofloaded to ES, jy,v = x + 1 represents the application Firstly, section 2 describes the system model and defines is ofloaded to cloud for processing. the problem formulation. Secondly, section 3 introduces our proposed multi-objective computation ofloading of VR application method. Then, section 4 demonstrates 2.1. Virtual Reality Application Model the experiment results. Finally, section 5 introduces the related work and section 6 describes the conclusion and the future work.
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 nodek represents the k−th node in application.
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 2.2.1. Executing latency and heterogeneous. These VR applications are executed by VDs or ofloaded 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 ⎧ ry,v lVe n,gwthhircehqiusedsetefindedbyastheVv-=th{avp1p,lvic2a,t.i.o.,nvovfny}-.tThhuesetarsiks Se(jy,v) = ⎪⎨ rfyfel,v denoted as ly,v, and the task workload is denoted as ry,v. ⎪⎩ rfyc,v Let X be the set of ES, denoted as X = {1, 2, ..., x}.
In each ES, the computing resources of ES are rep- 2.2.2. Transmission latency resented as a number of i virtual machine(VM) xn, represented as xn = {xn1, xn2, ..., xni}. J is a set of computation ofloading strategies, which is defined as J = {j1,1, ..., j1,v, j2,1, ..., jy,v}. And jy,v represents
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
tation ofloading strategy of the two tasks. Let P c(c ∈ {1, 2, 3}) be the flag between two strategies, which are defined as ⎧ T wo applications are executed ⎪ ⎪⎪⎪⎪ by same platf orm ⎪ P c = ⎪⎨⎪ V D and ES
of f loading ⎪⎪⎪⎪⎪ Cloud and ⎪ ⎪⎩⎪ other platf orms of f loading c = 3.
(2) According to transmission strategy P c, the transmission time St(jy,v) of the v-th application of the y-th VU is shown as
St(jy,v) = ly,v , (3)
B where B represents the transmission bandwidth. The transmission bandwidth is divided into three cases which is denoted as c = 1, c = 2, B = ⎪⎧⎨ B∞e ⎪⎩ Bc
P c = 1, P c = 2, P c = 3. 2.2.3. Waiting latency
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 xnix = (vmwk, tnum), where vmwk represents the ith VM collection and the tnum represents the number of applications in VM. When the ofloading strategy jy,v = x, ry,v is ofloaded to VM of x-th ES. Then, the vmwk is updated by vmwk = vmwk + ly,v and the tnum = tnum + 1. Finally, the Sw(jy,v) of v-th application from y-th VU is calculated as
Sw(jy,v) = Se(pre(jy,v)) · ψy,v,q, (5) 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 Se(pre(jy,v)). 2.2.4. Total latency
all time, including the executing latency Se(jy,v), the transmission latency St(jy,v) , and the waiting latency Sw(jy,v). The total latency is represented as S(jy,v) = Se(jy,v) + St(jy,v) + Sw(jy,v). (6)
2.3.1. Executing energy consumption
cuting time, the executing time of v-th VR application of y-th user is represented as ⎧ ry,v
fl · ℘active Se(jy,v) = ⎪⎨ rfye,v · ℘idle ⎩⎪ rfyc,v · ℘idle 2.3.2. Transmission 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
N t(jy,v) = St(jy,v) · ℘trans, where ℘trans represents the energy consumption of transmission. 2.3.3. Waiting energy consumption
N W (jy,v) = SW (jy,v) · ℘idle. (8) (9) 2.3.4. Total energy consumption
sented 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)
puting 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 ζv = ηv = 1, if ex is occupied 0, Otherwise. 1, if v existed in ex 0, Otherwise.
as U E(j), which is represented as
U E(j) =
ζv.
V v=1 τ represent the amount of occupied ESs which is calculated by τ = [α(jy,v) − τ (jy,v)]2). (16) (11) (12) (13) (15)
VR privacy conflicts need to be considered. In this paper, we address privacy conflicts by placing applications on diferent ESs to avoid privacy leakage [ 14]. A graph = (RW, CT ) is used to describe the privacy conflicts, where RW represents the a set of computing services and CT denotes a set of privacy conflicting relation. A pair of conflict relations (rwy, rwy∗)(rwy, rwy∗ ∈ RW ) are used to indicate that VU information with privacy conlficts cannot be placed in the same ES. The conflict application of rwv are represented as rwv|(rwy, rwy∗) ∈ CT , y∗ = {1, 2, ..., Y }. (17) HS = hs1, hs2, ..., hsy(hs ∈ Y ) represents the placement location ES for executing the applications. Afterwards, based on the acquired conflicting service set, the placed destination hsy has a conflicting ES set, which is calculated by hy|(esx esx) ∈ rwq, x = {1, 2, ..., |rwv|}. (18)
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.
M in M in M in
Y
V y=1 v=1 Y
V y=1 v=1 Y
V
The initialization of the algorithm includes the defi- The two populations Zd and Zd∗ generated by the nition of key parameters and the generation of the initial algorithm need to select the number of tz as the nextpopulation. Firstly, the first-generation population Z0 of generation population Zd+1, where d<Dmax. In addisize tz is generated randomly. Secondly, MCOVR defines tion, based on R2 ranking of Zd, by comparing the obsome parameters, such as crossover probability Lc, muta- jective function values of the two solutions and select tion probability Lm, the number of iterations Dmax, and the optimal value of the objective function as the parent current iteration index d. Finally, the algorithm defines a solution. size of tq achieve set Q0 that keeps the achieve solution.
Moreover, tournament selection, crossover, as well as 3.5. Method Overview mutation is executed on Z0 to generate new population Z0∗.
In the crossover operation, the algorithm randomly exchanges a certain point value of two ofloading strategies. Through crossover operation, the algorithm can obtain diversity solutions. In the mutation operation, mutations will slightly change certain values on the ofloading 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 Tchebychef 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.
R2i(A, W ) = − 1
min ϑ(a). |W | ω∈W a∈A
(25)
Additionally, in Equation (25), when each population is calculated by the R2 indicator, the reference point of the ϑ needs to be updated. The maximum value cmax and minimum value cdmin of the objective need to d be found. Through Equation (25) and reference point updating policy, the r2 ranking can be calculated as rankZd = (26) where fd(jy,v) represents the objective function in d-th iteration.
fd(jy,v) − cdmin We design three diferent kinds of comparative exω∈W a∈mAi/nBk{i∈m1,a..x.,t wi| cmax − cdmin |}, periments to verify the efectiveness of MCOVR. All the d 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 diference of the latency consumed among the three methods is getting larger. Our proposed
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.
three platforms for processing randomly, named as Benchmark.
First Come First Service (FCFS): All applications are ofloaded to the three platforms for processing orderly, named as FCFS.
The key parameter value is described in Table 1. We execute these ofloading 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.
MCOVR achieves the minimum latency consumption among the three methods. The latency consumption experimental result of the influence of diferent applications and diferent nodes are represented in Figure 4(b) and Figure 4(c). MCOVR reduces latency significantly and shows the slowest growth trend among the three methods. It can be deduced that MCOVR outperforms FCFS and Benchmark in terms of diferent situations.
The energy consumption is obtained by Equation (10). The energy consumption is related to time consumption which has the same growth trend with time consumption. Figure 5 demonstrates the comparison result of energy consumption. The multi-VD, multi-application and multi-node comparison of energy consumption is shown as Figure 5(a)-Figure 5(c). It is can be seen that MCOVR is more energy-eficient than the other two methods.
The load balancing can be obtained by Equation (16). It is a negative indicator, that means the low value of load balancing represents balanced tasks allocation in the ES cluster. The load balancing comparison result of three methods is shown in Figure 6. As shown in Figure 6(a) and Figure 6(b), with the number of tasks increases, MCOVR and FCFS show a downward trend and the diference between the two methods is little. Due to the ofloading strategies of Benchmark are generated randomly. Therefore, the results of Benchmark are unpredictable and irregular. As shown in Figure 6(c), as the number of tasks increases, more task are ofloaded to the ESs, and the load balancing value decreases of these three methods.
Above all, it is observed that our proposed method can obtain the optimal solution in comparison to the three experiments. ofloading 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 eficiency of VR application.
Diferent from the above studies, we investigates how to make privacy-aware computation ofloading 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 ofloading 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 efective.
In future work, we will study mobility-aware computation ofloading for augmented reality applications in smart cities.
Many eforts have been make to solve the computation ofloading in MEC. Bi et al. [ 9] considered energy consumption with the maximum number of memory constraint in MEC. An eficient 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 ofloading in the wireless powered MEC. Li et al. [18] studied a joint optimization problem of computation ofloading 7. Acknowledgments and privacy preservation in the MEC system and proposed a privacy-aware online learning algorithm to solve Acknowledgments this challenge. Ko et al. [19] studied the computation ofloading problem in smart cities based on the privacy This work is supported by the National Science entropy model. A constrained Markov decision process Foundation of China (Grant No.61902133), the Natmethod is developed to solve location privacy leakage. ural Science Foundation of Fujian Province (Grant The above-mentioned studies provided a good view of No.2018J05106), the Fundamental Research Funds for the the computation ofloading in MEC and inspired us to Central Universities(ZQN-817), Quanzhou Science and study the computation ofloading of VR applications. Technology Project(No.2020C050R).
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 References or cloud. They proposed a service placement algorithm for multi-user VR applications that makes placement de- [1] Mustafa, T., & Varol, A. (2020, June). Review of the Incisions, based on QoS and VUs mobility patterns. Zhu et ternet of Things for Healthcare Monitoring. In 2020 al. [21] focused on dynamic rendering-module placement 8th International Symposium on Digital Forensics problem for VR streaming. To reduce such rendering la- and Security (ISDFS) (pp. 1-6). IEEE. tency, they proposed a prediction-based pre-rendering [2] Wijethilaka, S., & Liyanage, M. (2021). Survey on mechanism at the ES. In [22], the authors investigated network slicing for Internet of Things realization the problem of providing service using the MEC network in 5G networks. IEEE Communications Surveys & architecture with a frame rendering. They solved an Tutorials, 23(2), 957-994. online service placement problem by leveraging model [3] Liu, J., & Zhang, Q. (2020). To improve service relipredictive control and overcoming conflicting system ability for AI-powered time-critical services using objectives. In [23], the authors focused on the online imperfect transmission in MEC: An experimental study. IEEE Internet of Things Journal, 7(10), 9357- works. IEEE Internet of Things Journal. 2021. 9371. [16] GÓmez, R. H., & Coello, C. A. C. (2013, June). [4] Chakareski, J. (2020). Viewport-Adaptive Scalable MOMBI: A new metaheuristic for many-objective Multi-User Virtual Reality Mobile-Edge Streaming. optimization based on the R2 indicator. In 2013 IEEE IEEE Transactions on Image Processing, 29, 6330- Congress on Evolutionary Computation (pp. 24886342. 2495). IEEE. [5] Hu, F., Deng, Y., Saad, W., Bennis, M., & Aghvami, A. [17] Feng, J., Pei, Q., Yu, F. R., Chu, X., & Shang, B. (2019).
H. (2020). Cellular-connected wireless virtual real- Computation ofloading and resource allocation for ity: Requirements, challenges, and solutions. IEEE wireless powered mobile edge computing with laCommunications Magazine, 58(5), 105-111. tency constraint. IEEE Wireless Communications [6] Hsu, Y. H., Cheng, J. H., Liao, K. Y., Wang, Y. S., Letters, 8(5), 1320-1323.
Chen, T. H., Chen, H. Y., ... & Liao, W. (2020, Septem- [18] Li, T., Liu, H., Liang, J., Zhang, H., Geng, L., & ber). NTU Smart Edge for Wireless Virtual Reality. Liu, Y. (2020, September). Privacy-Aware Online Task In 2020 IEEE International Conference on Consumer Ofloading for Mobile-Edge Computing. In InternaElectronics-Taiwan (ICCE-Taiwan) (pp. 1-2). IEEE. 1 tional Conference on Wireless Algorithms, Systems, [7] Sukhmani, S., Sadeghi, M., Erol-Kantarci, M., & El and Applications (pp. 244-255). Springer, Cham.
Saddik, A. (2018). Edge caching and computing in [19] Ko, H., Lee, H., Kim, T., & Pack, S. (2020). LPGA: 5G for mobile AR/VR and tactile internet. IEEE Mul- Location Privacy-Guaranteed Ofloading Algorithm tiMedia, 26(1), 21-30. in Cache-Enabled Edge Clouds. IEEE Transactions [8] Shi, S., Gupta, V., Hwang, M., & Jana, R. (2019, June). on Cloud Computing.
Mobile VR on edge cloud: a latency-driven design. In [20] Zhang, W., Chen, J., Zhang, Y., & Raychaudhuri, D. Proceedings of the 10th ACM Multimedia Systems (2017, October). Towards eficient edge cloud augConference (pp. 222-231). mentation for virtual reality mmogs. In Proceedings [9] Bi, J., Yuan, H., Duanmu, S., Zhou, M. C., & Abu- of the Second ACM/IEEE Symposium on Edge Comsorrah, A. (2020). Energy-optimized Partial Compu- puting (pp. 1-14). tation Ofloading in Mobile Edge Computing with [21] Zhu, Z., Jiang, N., Guo, T., & Wei, S. (2019, Genetic Simulated-annealing-based Particle Swarm November). Virtual reality streaming at the edge: Optimization. IEEE Internet of Things Journal. a power perspective: poster. In Proceedings of the [10] Xu, X., Gu, R., Dai, F., Qi, L., & Wan, S. (2020). Multi- 4th ACM/IEEE Symposium on Edge Computing (pp. objective computation ofloading for internet of ve- 316-318). hicles in cloud-edge computing. Wireless Networks, [22] Zhang, Y., Jiao, L., Yan, J., & Lin, X. (2019). Dynamic 26(3), 1611-1629. service placement for virtual reality group gaming [11] Violos, J., Psomakelis, E., Tserpes, K., Aisopos, F., on mobile edge cloudlets. IEEE Journal on Selected & Varvarigou, T. (2019, May). Leveraging user mo- Areas in Communications, 37(8), 1881-1897. bility and mobile app services behavior for optimal [23] Zhu, H., Li, Y., Chen, Z., & Song, L. (2021, March). edge resource utilization. In Proceedings of the In- Mobile Edge Resource optimization for Multiplayer ternational Conference on Omni-Layer Intelligent Interactive Virtual Reality Game. In 2021 IEEE WireSystems (pp. 7-12). less Communications and Networking Conference [12] Xu, X., Liu, X., Xu, Z., Wang, C., Wan, S., & Yang, (WCNC) (pp. 1-6). IEEE.
X. (2019). Joint optimization of resource utilization and load balance with privacy preservation for edge services in 5G networks. Mobile Networks and Applications, 1-12. [13] O’Brolcháin, F., Jacquemard, T., Monaghan, D.,
O’Connor, N., Novitzky, P., & Gordijn, B. (2016). The convergence of virtual reality and social networks: threats to privacy and autonomy. Science and engineering ethics, 22(1), 1-29. [14] Xu, Z., Gu, R., Huang, T., Xiang, H., Zhang, X., Qi, L., & Xu, X. (2018). An IoT-oriented ofloading method with privacy preservation for cloudlet-enabled wireless metropolitan area networks. Sensors, 18(9), 3030. [15] Chen, X., & Liu, G. Energy-Eficient Task Ofloading and Resource Allocation via Deep Reinforcement Learning for Augmented Reality in Mobile Edge Net