A Computation Offloading Scheme for UAV-Edge Cloud Computing Environments Considering Energy Consumption Fairness

Round 1

Reviewer 1 Report (New Reviewer)

- in the abstract, it suffices to write "about 120%", the decimals do not add to the statement
- line 164 is not clear: "The proposed algorithm in that is referenced in our research has problems that it does not consider the energy consumption and energy fairness of each node."
- 168: Should "genetic algorithm-based adaptive offloading" not abbreviate to GAAO?
- in line 219 you state that "we assume a UAV-Edge-Cloud computing environment in our energy consumption fairness-aware computation offloading scheme [which is] reasonable because it is similar to modern mobile computing environments." What are you trying to say here? Assuming X because it is similar to Y is not a good motivation (I hope this comment makes sense). You then basically argue that there are different levels of computational power available in different layers, and that some fast connection exists. But you do not model these or attribute cost to them.
- Your results (and your conclusion) say nothing about scalability. You tested three problem sizes, how does your approach fare for much smaller and much larger problems? What is the impact of swarm size over swarm distribution? What is the problem range for which your approach outperforms the one proposed by Liu?
- I am also missing the Future Work.


The manuscript certainly shows that a lot of work was undertaken. However, it seems that the work presented is merely an improvement over previous work of Liu [3]. I would like some comparison to other approaches. I know that this constitutes the dreaded "do more experiments" comment but I would settle for a well-informed discussion that compares the approach and the problem to other variations of the problem (and their proposed solutions) in the literature. A well-informed argument detailing the area within the field / solution-approach space where your approach is likely to perform well (or even better than others) would be appreciated. As it stands your work does not enable the reader to decide whether this approach is effectively a better choice for a specific problem the reader might want to solve themselves. If you offer such a critical discussion (well, or if you offer more comparative results) I would be willing to consider the manuscript as an accept.


A quick search gave me these (listing just a few) publications:

- https://www.sciencedirect.com/science/article/pii/S1084804522000108
A survey sketches the literature landscape and allows you to provide a better identification of where you research fits inside the larger field, as well as what the overall contributions of your work are and to which extend they move the field forward.

- https://ieeexplore.ieee.org/document/9685920
This article (e.g.,) which provides a far more detailed model. I would like to know the simplifications your model contains, and would like to have at least some model for your system model.

- https://ieeexplore.ieee.org/document/8641189
This article also provides a more detailed model. It also uses an entirely different approach. I would like to know what performance differences you would expect in comparison and some insight when one approach would be more suitable than the other. Surely yours does not outperform all others in all circumstances, right?

- I have found two more papers that do a better job at formalizing the problem / system:
    https://www.mdpi.com/1099-4300/24/5/736
    https://www.mdpi.com/2076-3417/12/13/6566
    

Author Response

We highly appreciate your valuable comments.

We revised our manuscript as your comments.

Please, find the attached file.

Author Response File: Author Response.pdf

Reviewer 2 Report (New Reviewer)

 This paper proposed a computational offloading and workflow allocation scheme while considering the energy consumption fairness for UAV-Edge-Cloud computing environments.  It was shown that the performance of the proposed scheme based on Genetic Algorithm was superior to that of Liu et al.’s Markov Approximation-based scheme in terms of the energy consumption fairness. I have the following comments.

1） Why does the author use GA algorithm?

2）In the actual scenario, the transmission power of  UAV is limited. Whether the transmission power of UAV should be a constraint  in the actual scenario. 

3）In 4.2.3. Figure 6 shows the simulation results of the average distance between the nodes that were allocated tasks. I think it should be Figure 7.

4) In 2.3, the author introduces Genetic Algorithm-based offloading schemes. In section 4, why did the author not compare the proposed algorithm with other GA based algorithms?

5）In reference [3], the Markov approximation-based algorithm (MA algorithm) was used to solve this NP-hard problem with the objective to minimize both the computation and routing cost. The proposed scheme was compared with Liu et al.’s scheme [3].  Whether the objective functions and constraints of the two optimization problems are consistent？Why did the author choose the algorithm proposed by Liu for comparison？

6) I suggest the authors give details and rationality of simulation parameters.

7）The energy consumption models were based on Huang et al. [26] and Dusza et al. [28]. Please give the detailed reasons for selecting energy consumption models.

Author Response

We highly appreciate your valuable comments.

We revised our manuscript as your comments.

Please, find the attached file.

Author Response File: Author Response.pdf

Reviewer 3 Report (New Reviewer)

Dear authors,

       The article, which you have prepared, is on the topic of the issue Special Issue "Edge Computing and IoT Technologies for Drones".

It is not clear from “Introduction” what scientific and technical problem is solved with UAV Swarms and the UAV-Edge-Cloud model computing environment. For comparison, the article by Liu et al. gives a clear Figure1 from which it follows that the problem is solved in the field of computer vision and "object recognition". If it is difficult for you to change the Figure 1, could you, please, write about it? Introduction contains a sufficient review of the sources on the chosen topic. 

In section “2. Related Works”, you provided links to the articles and conference proceedings.  Since the topic of this paper is interdisciplinary, could please, cite famous monographs in the field of IoT and Genetic Algorithms?

For example,

1. Andy King. Programming the Internet of Things: An Introduction to Building Integrated, Device-to-Cloud IoT Solutions. 2021. 421 p.

2. Dimitrios Serpanos, Marilyn Wolf. Internet-of-Things (IoT) Systems: Architectures, Algorithms, Methodologies. 2018, 179 p.

3.Dan Simon. EVOLUTIONARY OPTIMIZATION ALGORITHMS. Biologically-Inspired and Population-Based Approaches to Computer Intelligence. 2013. 784 p.

4. Eyal Wirsansky. Hands-On Genetic Algorithms with Python: Applying genetic algorithms to solve real-world deep learning and artificial intelligence problems  2020. 348 p.

In section 3.2. “Proposed Computation Offloading Algorithm” what is missing is a complete mathematical statement of the problems. From a mathematical point of view the similar in meaning articles by Liu et al. [3] and by Yu et al. [8] are more neatly written. They had a problem statement, the optimization problem was formulated, the cost function was introduced, and theorems,  lemmas were formulated.

In your paper you are trying to improve the results of the energy consumption method by highly cited Liu et al. paper. By comparison, your paper uses only 3 formulas, while Liu et al. paper has 48 formulas, Yu et al. has 43 formulas. Please, improve this part to make it clear that you are solving an optimization problem.

In section “4.1. Simulation Environments”, there are no any technical information about nodes (drones, GPU computing servers, cloud servers).

What are the values of different processing power, network bandwidth, and latency for these nodes? What are the values of Bandwidth of links in Mb/s between UAV to UAV, UAV to Edge, Edge to Edge, Edge to Cloud?

What are the maximum operating time for drones?  Could you, please, provide this data in a separate table?  Please, see Liu et al. [3], Table II.

You also wrote: “We assume that an edge server is equipped with multiple GPUs.  Therefore, an edge server has more powerful computing power when compared to each UAV. It is assumed that a cloud server has a better computing performance than an edge server. On the other side, edge and cloud servers have more considerable network delays than UAVs, but they can provide more powerful computing capabilities compared to each UAV.”

Please, give specific data about your cloud server that "has a better computing performance than an edge server with multiple GPUs"? What is the model of such cloud server?

These days you can easily buy a 3U server with 2 Intel Xeon CPUs, 2 TB RAM and 8 Nvidia A100 GPU cards for machine learning tasks. About what models of GPUs are you talking in your case? 

There are a lot of general phrases in your article. Could you, please, give specific characteristics of each node?

In section “4. Performance Evaluations 4.1. Simulation Environments” , please, justify why you have selected only 2 cloud servers. Usually there are dozens or hundreds of servers in the cloud.  Do you agree that two servers in the cloud is clearly not enough?

How will your algorithm work if you increase the number of servers in the cloud, for example, from 10 to 100?

In your simulation, each experiment was performed 100 times, and the results were averaged. Please, justify why you selected this amount.

The choice of "Scheme-dependent Parameters" values in Table 2 should also at least be justified.

In section “4.2. Simulation Results and Discussions” you showed on Figures 6,7 the simulation results of the energy consumption fairness and simulation results of the average distance between nodes. Your algorithm calculated "Fnew" which denotes the energy consumption fairness value of the generated new chromosome.

Could you present a graph of how the "Fnew" value behaves as a function of  Number of iterations (GA-based), Number of iterations (MA)? It would be correct to use graphs to demonstrate the convergence of your algorithm.

The specialists would be interested to read more information on how you have implemented your GA algorithm programmatically.

What programming language,  scientific libraries did you use? How long in time one typical case was run on you server or notebook?

In “5. Сonclusions” it makes sense to indicate how much % your algorithm gives a gain. The article can be published after revision.

Best regards,

Reviewer

Author Response

We highly appreciate your valuable comments.

We revised our manuscript as your comments.

Please, find the attached file.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report (New Reviewer)

I am still not convinced that this is of large interest to the community, but I no longer object to the paper being accepted. Congratulations.

Reviewer 2 Report (New Reviewer)

The authors answered the questions. I am satisfied with the revised version.

Reviewer 3 Report (New Reviewer)

Dear authors,

Thank you for taking all my comments into account and for substantially revising the article. I recommended the article for publication.

Best regards,

Reviewer

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

- It seems to be a "Travelling Salesman Problem", a problem that has been widely studied. Is this article really an adaptation of this problem to information processing in UAVs?

- If the answer to the previous question is yes, why has it not been compared with other solutions used to solve this problem?

- It is not specified why the use of GA is better than the use of Markov approximation, it would be necessary to make an estimation and explanation of why GA improves the results. It seems that an empirical demonstration is not conclusive of one system being better than another.

- The random process of chromosome permutations should be improved by means of graphs and diagrams, the explanation in the text is a little confusing.

Author Response

Responses for Reviewer #1

 

(Comment 1) It seems to be a "Travelling Salesman Problem", a problem that has been widely studied. Is this article really an adaptation of this problem to information processing in UAVs?

(Response)

We highly appreciate your careful comment. In this paper, it is not simply a matter of allocating tasks, but it is dealing with the problem of task allocation considering the energy consumption, and each task has logical precedence relationships. We focus on fairness issues to maximize the lifetime of the entire network consisting of multiple drones and various servers by using our energy consumption fairness model.

 

(Comment 2)  If the answer to the previous question is yes, why has it not been compared with other solutions used to solve this problem?

(Response)

We totally understand your concern. Our GA-based proposed scheme was compared with the MA-based scheme proposed by Liu et al.  However, existing Liu's allocation scheme did not consider the aspect of energy consumption fairness when allocating tasks. In our GA-based proposed scheme, energy consumption fairness was considered because the problem of imbalanced energy consumption of each drone is very important. We think that there is a difference in that we deal with energy fairness issues to maximize the lifetime of the entire network consisting of multiple drones and servers.

 

(Comment 3) It is not specified why the use of GA is better than the use of Markov approximation, it would be necessary to make an estimation and explanation of why GA improves the results. It seems that an empirical demonstration is not conclusive of one system being better than another.

(Response)

Thank you for the suggestions. It is a great point. According to your comments, the reason why the proposed GA-based scheme performed better than the MA-based scheme has been explained.  in more detail. Please see “Performance Evaluations” section in the revised manuscript.  “Overall, the proposed GA-based scheme shows better performance than the MA-based scheme in terms of the energy consumption fairness. These results are because the proposed GA-based scheme can explore more diverse allocation solutions in the problem space compared to the MA-based scheme. Therefore, the GA-based proposed scheme can find an allocation solution that is closer to the optimal global solution. In the case of a large network rather than a small network, this tendency was larger, and the GA-based proposed scheme showed higher performance in terms of the energy consumption fairness.”

 

(Comment 4)  The random process of chromosome permutations should be improved by means of graphs and diagrams, the explanation in the text is a little confusing.

(Response)

Thank you for your good comments. We have revised the manuscript by adding an example of chromosomes in the proposed algorithm and its explanation. The figure and the revised paragraph explain how a chromosome is composed of. Please see “Proposed Computation Offloading Algorithm” section in the revised manuscript. (In addition, we also revised several sentences and words about chromosomes and permutations in overall descriptions.) “Each workflow allocation information is combined to create a chromosome. Each chromosome’s information can be considered one of the candidates that can allocate given workflows onto the network topology. Figure 2 shows an example of three chromosomes for a given network topology and two workflows. In Figure 2, w1 consists of three tasks, t1, t2, and t3, w2 consists of two tasks, t4 and t5. Allocw1 and Allocw2 represent the allocated nodes of the tasks in w1 and w2, respectively. And a chromosome is a concatenation of Allocw1 and Allocw2. For example, the chromosome is represented as ((3, 7, 5), (1, 2)) in the case of Figure 2(a).”

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors propose the optimization mechanism to minimize the energy consumption of a drone cluster. Nevertheless, the method they propose is old, well known and established method, therefore I do not see anything original in the paper. Moreover, it is also known that 2-opt method used here can  tend to stuck in a local optimization minimum. To resolve this, better methods were proposed already in 1972, so called V-opt method. In k-opt methods the chromosome is cut and recombined at fixed number of points (k) while in v-opt methods the number of points and recombination between parents is varied at each step. In this way the process in less likely to stuck in a local minimum.

In addition, the algorithm is tested (simulated) only on two cases, which is way to little to prove the statement and is compared with the MA which is known to be one of the worst performing optimization methods anyway.

Author Response

Responses for Reviewer #2

 

(Comment 1) The authors propose the optimization mechanism to minimize the energy consumption of a drone cluster. Nevertheless, the method they propose is old, well known and established method, therefore I do not see anything original in the paper. Moreover, it is also known that 2-opt method used here can tend to stuck in a local optimization minimum. To resolve this, better methods were proposed already in 1972, so called V-opt method. In k-opt methods the chromosome is cut and recombined at fixed number of points (k) while in v-opt methods the number of points and recombination between parents is varied at each step. In this way the process in less likely to stuck in a local minimum. 

(Response)

Thank you for your good comments.  We applied the GA-based scheme to search and find better allocation solutions in the problem space without being stuck in a local minimum. To avoid being stuck in a local minimum, we applied a one-point crossover operation and mutation operation. The goal of our GA-based workflow allocation scheme is to find an allocation solution for given workloads that can maximize the energy consumption fairness based on GA. It is not the scope of our research to improve GA performance itself.

 

(Comment 2) In addition, the algorithm is tested (simulated) only on two cases, which is way to little to prove the statement and is compared with the MA which is known to be one of the worst performing optimization methods anyway. 

(Response)

Thank you for your good comments.  According to your comments, we have performed additional simulations to evaluate the performance of the proposed scheme. The performance of the proposed scheme was compared and analyzed in the network topologies: small(50×50), medium(100×100), and large(200×200). And, we have added the results to the "Performance Evaluations" section. Please see “Performance Evaluations” section in the revised manuscript.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper tackles the problem of allocating computing tasks among a set of heterogeneous devices with limited resources, as is the case of Drones. Authors propose to tackle this challenge by means of a framework able of allocating and offloading tasks evenly among drones while being aware of energy consumption in each drone. The propose decision method is based on a Genetic Algorithm (GA). This proposal is motivated by the fact that computational offloading schemes can be applied to reduce the energy consumption of drones.

The paper reads well, but authors may need to revise it to fix some small issues. For instance on section 3.1 the sentence "W means the set of all workflows to be allocated." has no meaning since the reader cannot connect the variable W with the proposed system, since the reference to table 1 comes later on in the paper.

Most of the offloading architectures presented/analyzed in the paper are suitable for IoT systems, while the described goals of the paper seem to be more generic. So the question is: why did not the authors analyse other architectures besides IoT? Nevertheless, authors present a good analysis of offloading algorithms, going from SDN based on decentralized algorithms that allow drones to take decisions about offloading tasks.

Authors propose a scheme that aims to evenly distribute computational effort to all drones. However this proposal seems to assume that all drones have the same level of energy capacity (homogeneous), while in reality different drones may have different energy capacity.

Moreover, authors did not provide a justification of why their approach is better than other approaches that also aim to support computing offloading tasks based on Genetic Algorithms, such as "Collaborative Task Offloading Strategy of UAV Cluster Using Improved Genetic Algorithm in Mobile Edge Computing" (2021) or "Adaptive offloading in mobile-edge computing for ultra-dense cellular networks based on genetic algorithm". Simulation results are compared with a Markov Approximation based scheme, providing no evidence that this proposal performs better than other GA based approaches. Moreover results related with the total energy consumption seem to point to the fact that the proposed solutions lead to a higher energy consumption than MA in large networks.

Hence, it is advised that authors provide a better analysis of other GA related work, in order to show the novelty of the proposed approach. As for the experimental results, a comparison with other GA mechanisms would be desirable, and a justification of the poorer performance in large networks is needed. 

Author Response

 Responses for Reviewer #3

 

(Comment 1) The paper reads well, but authors may need to revise it to fix some small issues. For instance on section 3.1 the sentence "W means the set of all workflows to be allocated." has no meaning since the reader cannot connect the variable W with the proposed system, since the reference to table 1 comes later on in the paper.

(Response)

Thank you for your good comments.  As the reviewer mentioned, we have changed the location of Table 1 and the location of the reference to Table 1.

 

(Comment 2) Most of the offloading architectures presented/analyzed in the paper are suitable for IoT systems, while the described goals of the paper seem to be more generic. So the question is: why did not the authors analyse other architectures besides IoT? Nevertheless, authors present a good analysis of offloading algorithms, going from SDN based on decentralized algorithms that allow drones to take decisions about offloading tasks.

(Response)

We highly appreciate your careful comment.  In response to the reviewer's comments, we explain the main contributions of our paper as follows: The main contributions of this paper are as follows: - Our system model has three layers: UAV-based computing area, edge computing area, and cloud computing area. We consider the routing cost among drones for each offloading task in the UAV-based computing area. We also consider the scheduling cost of requests of multi-users in the edge and the cloud computing area. - The energy efficiency is an important issue in drone-based systems, but the energy consumption fairness is also an important issue to sustain the network connection among UAVs. The originality of our GA-based computation offloading and workflow allocation scheme considers the residual energy balance among UAVs to extend the lifetime of the UAV-based computing area. - We use a GA-based computation offloading decision scheme to find a better solution by considering the energy consumption, the energy consumption fairness, and the resource constraints such as processing power or network bandwidth. Our approach is outperformed compared with the previous Markov approximation-based scheme in terms of the energy consumption fairness.

 

(Comment 3) Authors propose a scheme that aims to evenly distribute computational effort to all drones. However this proposal seems to assume that all drones have the same level of energy capacity (homogeneous), while in reality different drones may have different energy capacity.

(Response)

Thank you for your good comments.  The proposed scheme also considers the case where each drone's energy capacity differs. The proposed scheme initially distributes tasks of each workflow to UAV-Edge-Cloud computing environments and distributes the tasks considering the energy consumption of each. Therefore, if the remaining energy of a certain drone is not enough to handle the task, the task is not allocated to the drone. Thus, the proposed scheme considers the situation where the energy capacity is not only homogeneous but also heterogeneous. We have revised the manuscript to clarify this issue as follows. Please see the “Proposed Computation Offloading Algorithm” sub-section. “If no node can replace vj because of resource constraints (such as processing power, network bandwidth, or remaining energy) or transmission range, the assignment information is not changed.”

 

(Comment 4) Moreover, authors did not provide a justification of why their approach is better than other approaches that also aim to support computing offloading tasks based on Genetic Algorithms, such as "Collaborative Task Offloading Strategy of UAV Cluster Using Improved Genetic Algorithm in Mobile Edge Computing" (2021) or "Adaptive offloading in mobile-edge computing for ultra-dense cellular networks based on genetic algorithm". Simulation results are compared with a Markov Approximation based scheme, providing no evidence that this proposal performs better than other GA based approaches. Moreover results related with the total energy consumption seem to point to the fact that the proposed solutions lead to a higher energy consumption than MA in large networks.

(Response)

Thank you for your good comments.  According to your comments, genetic algorithm-based offloading schemes have been added to the “Related works” section, and differences from the related studies have been added. The related works you mentioned and our paper are different from each other. In our paper, we assume that a workflow consists of multiple tasks. We assume that streaming data are constantly being transferred and processed between tasks. Each task has logical precedence. In this situation, the goal of our GA-based workflow allocation scheme is to find an allocation solution for given workloads that can maximize the energy consumption fairness based on GA. And the reason why the proposed GA-based scheme performed better than the MA-based scheme has been explained in more detail in the “Performance Evaluations” section as follows. “Overall, the proposed GA-based scheme shows better performance than the MA-based scheme in terms of the energy consumption fairness. These results are because the proposed GA-based scheme can explore more diverse allocation solutions in the problem space compared to the MA-based scheme. Therefore, the GA-based proposed scheme can find an allocation solution that is closer to the optimal global solution. In the case of a large network rather than a small network, this tendency was larger, and the GA-based proposed scheme showed higher performance in terms of the energy consumption fairness.” We also attached a list of references on  GA-based offloading schemes as follows: 18. Hussain, A.; Manikanthan, S.V.; Padmapriya, T.; Nagalingam, M. Genetic algorithm based adaptive offloading for improving IoT device communication efficiency. Wireless Networks 2019, 26, 2329-2338. https://doi.org/10.1007/s11276-019-02121-4. 19. Liao, Z.; Peng, J.; Xiong, B.; Huang, J. Adaptive offloading in mobile-edge computing for ultra-dense cellular networks based on genetic algorithm. Journal of Cloud Computing 2021, 10, 1-16. https://doi.org/10.1186/s13677-021-00232-y. 20. Wang, H. Collaborative Task Offloading Strategy of UAV Cluster Using Improved Genetic Algorithm in Mobile Edge Computing. Journal of Robotics 2021, 2021, 1-10. https://doi.org/10.1155/2021/3965689. 21. Li, Z.; Zhu, Q. Genetic Algorithm-Based Optimization of Offloading and Resource Allocation in Mobile-Edge Computing. Information 2020, 11. https://doi.org/10.3390/info11020083. 22. Chen, Z.; Zheng, H.; Zhang, J.; Zheng, X.; Rong, C. Joint computation offloading and deployment optimization in multi-UAV-enabled MEC systems. Peer-to-Peer Networking and Applications 2022, 15, 194-205. https://doi.org/10.1007/s12083-021-01245-9. 471 23. Chakraborty, S.; Mazumdar, K. Sustainable task offloading decision using genetic algorithm in sensor mobile edge computing. Journal of King Saud University - Computer and Information Sciences 2022, 34, 1552-1568. https://doi.org/https://doi.org/10.1016/j.jksuci.2022.02.014.

 

(Comment 5) Hence, it is advised that authors provide a better analysis of other GA related work, in order to show the novelty of the proposed approach. As for the experimental results, a comparison with other GA mechanisms would be desirable, and a justification of the poorer performance in large networks is needed.

(Response)

Thank you for your good comments.  According to your comments, genetic algorithm-based offloading schemes have been added to the “Related works” section, and differences from the related studies have been added. According to your comments, we have performed additional simulations to evaluate the performance of the proposed scheme. The performance of the proposed scheme was compared and analyzed in the network topologies: small(50×50), medium(100×100), and large(200×200). And we have added the results to the "Performance Evaluations" section.  In addition, we have added explanations for the poorer performance in the large network topology in the “Performance Evaluations” section as follows. “In the case of the proposed GA-based scheme, the energy consumption fairness is reduced in the case of the large network topology than in the case of the small network topology. This result is due to the different values of simulation parameters applied to each network topology. In the case of the large network topology, the number of workloads to be allocated is larger than that of the small network topology, as shown in Table 2. The average number of tasks allocated to each drone increases, and the degree of imbalance in terms of the energy consumption of each drone may be greater than that of the small network topology. In addition, as the size of the network topology increases and the number of nodes in the network topology increases, the size of the problem space the proposed GA-based scheme explores also increases. Therefore, more various allocation solutions are possible, and it is more difficult to find the global optimum than the small network topology. For these reasons, the energy consumption fairness was slightly reduced.”

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

I still think that the manuscript does not present anything original. The proposed method in an old and already known method and its application on drones is straightforward and does not impose anything original.  

Reviewer 3 Report

The paper still has some limitations to show why the proposed solution is better. First, because no evaluation is done against alternative solutions. Evaluation is performed against MA, and in this case results show that the proposed GA scheme leads to higher energy consumption than MA in large and medium size networks. So, based on the presented evaluation results it is not possible to judge about the novelty that this proposal may have in relation to alternative solutions.