An Improved Genetic Algorithm with Swarm Intelligence for Security-Aware Task Scheduling in Hybrid Clouds

Round 1

Reviewer 1 Report

An interesting paper that focuses on the task scheduling problem with deadline and security constraints in environments with public & hybrid clouds via AI so as to increase user satisfactiom without performanc costs. The authors compare well their approach with a multitude of existing ones in the literature. 

Overall, the paper is matur and with contribution. Acceptance is suggested provided that the authors improve the quality of the discussion of the findings in the performance evaluation section, as these seem a bit weak

Author Response

Thanks for recognizing our contributions, and we have added more discussions on experiment results.

Reviewer 2 Report

The paper focuses on the task allocation problem in a hybrid cloud computing environment. The main contribution are i) an MINLP model definition, ii) the design of a heuristic algorithm that merges together PSO and GA approaches, and iii) an evaluation of the proposed algorithm in comparison to others that were performed based on an abstract model of hybrid cloud.  

The main drawbacks are:

1. The paper's novelty is limited because the considered problem has been studied so many times also considering much more constraints and more realistic conditions. Anyway, the proposed heuristic may be interesting if evaluated for practical, realistic cases.

2. There are no security issues considered at all. I my opinion the title is misleading as "security-aware task scheduling" is far more than just allocation in the private or public cloud. Please revise the title to be more related to the paper's content. 

3. I also have several doubts about the model. The first point is that orchestration is a process where new tasks arrive in the system and finished tasks leave it as well as other events happen. The main orchestration challenge is to deal with a such dynamic, uncertain scenario.  So, there is no analysis of a practical use case and no results in a realistic scenario. The second point, eq. 9 assumes that data transfer time per task is ai/rj (or ai/bk) per ith task, which is not true because data of several tasks processed on the same VM  may or may not be transferred simultaneously. This seems an error in the model, please clarify. Third point, the assumptions about only one core per task (eq. 3), non-work-conserving behavior, and treating idle periods as busy in eq. 20, as well as no weights in the objective function eq. 21 are quite strange assumptions requiring strong justification. 

4. The evaluation is performed on an artificial model of a hybrid cloud, where most of the data were randomly selected. The key question is this model representative and for what hybrid cloud? And the next questions are: Are randomly generated data consistent (i.e. all tasks are feasible)? Are presented conclusions justified? How assumed values influenced conclusions? What will happen if someone considers another model of hybrid cloud?  Moreover, there is no comparison of the proposed heuristic with the proposed MINLP model. I recommend including an example of how far is heuristic from the optimum. 

5. I have doubts if the paper's topic is reasonable for the Electronics journal. Please consider resubmission elsewhere.

6. Some minor issues: eq. 11 & 12 are not true in a general case, only for tasks assigned to the same core. bk is overdeclared. 

Author Response

[Comment 1] The paper's novelty is limited because the considered problem has been studied so many times also considering much more constraints and more realistic conditions. Anyway, the proposed heuristic may be interesting if evaluated for practical, realistic cases.

[Response 1] Thanks for the comment. In this paper, we focused on the task scheduling problem for hybrid cloud with security and deadline constraints, addressed by proposing a hybrid meta-heuristic. There are several works on the task scheduling with constraints of security or deadline for hybrid clouds, and some works exploiting hybrid meta-heuristics for the task scheduling problem in various distributed computing. But few works concern both.

 

[Comment 2] There are no security issues considered at all. I my opinion the title is misleading as "security-aware task scheduling" is far more than just allocation in the private or public cloud. Please revise the title to be more related to the paper's content.

[Response 2] Sorry for the confusion. We consider the two-level security model for the security constraints of tasks, where task with security requirement can be only processed by the private cloud, while other tasks can be processed by both private and public clouds. This is guaranteed by setting the possible cores for each task assignment during the whole procedure of our algorithm.

 

[Comment 3] I also have several doubts about the model. The first point is that orchestration is a process where new tasks arrive in the system and finished tasks leave it as well as other events happen. The main orchestration challenge is to deal with a such dynamic, uncertain scenario.  So, there is no analysis of a practical use case and no results in a realistic scenario. The second point, eq. 9 assumes that data transfer time per task is ai/rj (or ai/bk) per ith task, which is not true because data of several tasks processed on the same VM may or may not be transferred simultaneously. This seems an error in the model, please clarify. Third point, the assumptions about only one core per task (eq. 3), non-work-conserving behavior, and treating idle periods as busy in eq. 20, as well as no weights in the objective function eq. 21 are quite strange assumptions requiring strong justification.

[Response 3] For the first point, there are mainly two kinds of task scheduling problem, real-time scheduling and batch scheduling. The first kind of scheduling requires a sub-second latency of the scheduling decision for each new task. The second kind of scheduling can tolerate seconds or even minutes sometimes to make scheduling decisions for a batch tasks. In this paper, we focused on the batch tasks, which is a very common case in real world, such as delay schedulers, scheduling of high performance computing applications. For the second point, we use the sequential transmission mode for tasks, where there is only one task’s data transmission for each server in each time. This is guaranteed by Eq. (11) or (13), that there is no overlap between data transmissions of any two tasks in each server. For the third point, in this paper, we focused on the scheduling of single-thread tasks, and tasks with multiple threads can be seen as multiple single-thread tasks. In some cases, the network resources may be scarcer than the computing resources, especially for big data applications. Then some tasks’ computing must wait for the complete their data transmissions, leading to some cores are idle. For the objective function (21), the utilization (U) is no greater than 1. Thus, Eq. (21) means that the number of accepted tasks is the major optimization objective, and the utilization is the minor one.

 

[Comment 4] The evaluation is performed on an artificial model of a hybrid cloud, where most of the data were randomly selected. The key question is this model representative and for what hybrid cloud? And the next questions are: Are randomly generated data consistent (i.e. all tasks are feasible)? Are presented conclusions justified? How assumed values influenced conclusions? What will happen if someone considers another model of hybrid cloud? Moreover, there is no comparison of the proposed heuristic with the proposed MINLP model. I recommend including an example of how far is heuristic from the optimum.

[Response 4] In simulated experiments, parameters are set referring to related works. In addition, the range of tasks’ size are large, which can cover most of cases. To ensure the reliability of experimental results, we repeat our experiments more than 100 times, which means more than 100 simulated hybrid cloud environments are generated for our experiments.

 

[Comment 5] I have doubts if the paper's topic is reasonable for the Electronics journal. Please consider resubmission elsewhere.

[Response 5] The scope of the Electronics journal includes “Computer Science & Engineering” and “Systems & Control Engineering”, and both can cover the scope of our paper.

 

[Comment 6] Some minor issues: eq. 11 & 12 are not true in a general case, only for tasks assigned to the same core. bk is overdeclared.

[Response 6] Thanks for the comment. Constraints 11 & 12 are transformed into Eq. (13) and (14), where sum{xi1·xi2 + yi1·yi2} represents that two tasks assigned to the same cores (xi1, xi2 or yi1, yi2 are simultaneously 1) must satisfy Eq. (13) and (14). As illustrated in the main text, “Then, when tasks ti1 and ti2 are assigned to one core, we have Eq. (11) and (12), which

can be reformulated as Eq. (13) and (14) for all cases, respectively.”.

Reviewer 3 Report

The paper proposes a mixed-integer non-linear programming problem and develops a novel algorithm called SPGA to solve this problem. The proposed algorithm combines the self- and social cognitions of particle swarm optimization and the population evolution of genetic algorithms to optimize task scheduling with deadline and security constraints in a hybrid cloud.

The authors have also conducted extensive experiments to evaluate the performance of SPGA and compared it with 12 other scheduling algorithms. The results show that SPGA outperforms the other algorithms with higher accepted ratio and resource utilization.

Overall, the paper seems to be well-structured and presents a clear problem statement and proposed solution. The authors have also provided a comprehensive evaluation of the proposed algorithm.

 

Introduction: The introduction provides a clear motivation for the problem being addressed in the paper, which is the task scheduling problem in a hybrid cloud environment. The authors have highlighted the limitations of using a private cloud in meeting user demand and the benefits of using a hybrid cloud model. The introduction also explains the challenges involved in solving the task scheduling problem in a hybrid cloud environment and the existing approaches to address it.

The authors have provided a brief overview of the two categories of algorithms, heuristics and meta-heuristics, and how they are used to solve the task scheduling problem. They have also pointed out the strengths and weaknesses of these algorithms and the need for designing hybrid heuristic algorithms to achieve better performance.

Overall, the introduction is well-written and provides a clear overview of the problem being addressed and the research motivation. The authors have also provided a good background of the existing research in the area of task scheduling in a hybrid cloud environment. However, it would be useful if the introduction provided more details on the specific research questions that the paper aims to address and the contribution of the proposed algorithm.

 

Problem Statement: The task scheduling problem in hybrid clouds is formulated as an optimization problem that aims to maximize user satisfaction and resource efficiency while taking into consideration the deadline and security requirements of user request tasks. The problem involves assigning each requested task to a private/local resource based on a scheduling strategy. When private resources are insufficient, the provider rents resources from the public cloud. Tasks with security requirements can only be processed by private resources, as public resources cannot guarantee security due to being shared by various internet users. In cases where neither private nor public resources can meet a task's requirements, the request is rejected. The objective is to determine the optimal resource allocation for each task to maximize user satisfaction and resource efficiency.

 

Algorithm 1, called SPGA (Security-aware hybrid PSO and GA scheduling), is a high-level algorithm for task scheduling in a hybrid cloud environment. It takes as input the information of request tasks and available hybrid cloud resources and provides a task scheduling solution as output.

The algorithm starts by initializing a population of individuals (i.e., potential solutions), evaluating the fitness of each individual, and initializing each personal best and global best code. The algorithm then enters a loop that performs several operations on the population.

Firstly, the algorithm crosses each individual with another individual, its personal best and the global best codes, respectively, with the crossover probability. Then it evaluates the fitness of each offspring and replaces each individual with its best offspring. The algorithm updates each personal best and global best codes and mutates each individual with the mutation probability. It also evaluates the fitness of mutated individuals and updates each personal best and global best codes.

Finally, the algorithm decodes the global best code into a task scheduling solution using Algorithm 2 (not shown in this excerpt). The task scheduling solution is then returned.

Overall, the algorithm uses a combination of Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) techniques to optimize the task scheduling problem. The algorithm takes into consideration security requirements of the tasks, which are only processed by private resources in the hybrid cloud. However, without further information about the implementation and performance of the algorithm, it is difficult to assess its effectiveness.

Here are some issues that can be identified in the SPGA algorithm:

Lack of details: The algorithm lacks details about how the population is initialized and how the fitness of each individual is evaluated. Without these details, it's difficult to understand how the algorithm works and how effective it might be.

Crossover probability not defined: The algorithm uses crossover to combine genetic material from two individuals. However, the algorithm does not define the probability of performing a crossover operation, which can affect the exploration and exploitation capabilities of the algorithm.

Mutation probability not defined: The algorithm uses mutation to introduce new genetic material into the population. However, the algorithm does not define the probability of performing a mutation operation, which can also affect the exploration and exploitation capabilities of the algorithm.

Lack of explanation of personal best and global best codes: The algorithm mentions the use of personal best and global best codes, but does not explain how they are computed or updated. It is unclear how these codes relate to the task scheduling solution.

No explanation of how to handle security requirements: The algorithm mentions that some tasks have security requirements that can only be met by private resources. However, it does not provide any information on how to handle these tasks or ensure that they are assigned to private resources.

Lack of error handling: The algorithm does not account for any errors or exceptions that may arise during the scheduling process. It is important to include robust error-handling mechanisms to ensure the algorithm does not fail or produce incorrect

Cite the following paper as well:

Muhammad Sohaib Ajmal, Zeshan Iqbal, Farrukh Zeeshan Khan, Muneer Ahmad, Iftikhar Ahmad, Brij B. Gupta, Hybrid ant genetic algorithm for efficient task scheduling in cloud data centers, Computers and Electrical Engineering, Volume 95, 2021, 107419, ISSN 0045-7906, https://doi.org/10.1016/j.compeleceng.2021.107419

 

Author Response

[Comment 1] However, it would be useful if the introduction provided more details on the specific research questions that the paper aims to address and the contribution of the proposed algorithm.

[Response 1] Thanks for the comment. We have improve the introduction by highlight the problem we concerned in this paper.

 

[Comment 2] However, without further information about the implementation and performance of the algorithm, it is difficult to assess its effectiveness.

[Response 2] Thanks for the comment. We have improved the presentation of our algorithm.

 

[Comment 3] Here are some issues that can be identified in the SPGA algorithm:

Lack of details: The algorithm lacks details about how the population is initialized and how the fitness of each individual is evaluated. Without these details, it's difficult to understand how the algorithm works and how effective it might be.

[Response 3] Thanks for the comment. We have added more explanations on our algorithm.

 

[Comment 4] Crossover probability not defined: The algorithm uses crossover to combine genetic material from two individuals. However, the algorithm does not define the probability of performing a crossover operation, which can affect the exploration and exploitation capabilities of the algorithm.

Mutation probability not defined: The algorithm uses mutation to introduce new genetic material into the population. However, the algorithm does not define the probability of performing a mutation operation, which can also affect the exploration and exploitation capabilities of the algorithm.

[Response 4] In this paper, we focused on the integration method of hybrid meta-heuristics. We set parameters (including crossover and mutation operators) of SPGA simply referring to existing works, and their values are same to other methods in our experiments. This can verify the performance superiority objectively. Intuitively, parameters have impacts on the performance of PGA. Such impact will be studied experimentally in our future work.

 

[Comment 5] Lack of explanation of personal best and global best codes: The algorithm mentions the use of personal best and global best codes, but does not explain how they are computed or updated. It is unclear how these codes relate to the task scheduling solution.

[Response 5] Thanks for the comment. We have added more explanation on the personal best and global best codes.

 

[Comment 6] No explanation of how to handle security requirements: The algorithm mentions that some tasks have security requirements that can only be met by private resources. However, it does not provide any information on how to handle these tasks or ensure that they are assigned to private resources.

[Response 6] Sorry for the confusion. We consider the two-level security model for the security constraints of tasks, where task with security requirement can be only processed by the private cloud, while other tasks can be processed by both private and public clouds. This is guaranteed by setting the possible cores for each task assignment during the whole procedure of our algorithm.

 

[Comment 7] Lack of error handling: The algorithm does not account for any errors or exceptions that may arise during the scheduling process. It is important to include robust error-handling mechanisms to ensure the algorithm does not fail or produce incorrect

[Response 7] Thanks for the comment. We have added several error handlings in our algorithm.

 

[Comment 8] Cite the following paper as well:

 

Muhammad Sohaib Ajmal, Zeshan Iqbal, Farrukh Zeeshan Khan, Muneer Ahmad, Iftikhar Ahmad, Brij B. Gupta, Hybrid ant genetic algorithm for efficient task scheduling in cloud data centers, Computers and Electrical Engineering, Volume 95, 2021, 107419, ISSN 0045-7906, https://doi.org/10.1016/j.compeleceng.2021.107419

[Response 8] We have included this paper in the revised manuscript.