Multi-Task Scheduling Based on Classification in Mobile Edge Computing

Round 1

Reviewer 1 Report

This is an interesting work, but it seems misplaced in the context of IoT/Mobile-Edge. 

More specifically, mobile edge is about moving computation closer to the cellular subscriber, not about distributing the computational load between multiple mobile devices. See, for example, https://en.wikipedia.org/wiki/Mobile_edge_computing 

This work could fit in a context of IoT devices in the home where there may be some rationale to distribute complex computational tasks across multiple devices. However, the experiments provided in this case, based on MNIST and Imagenet, are not well matched with an IoT scenario. 

The underlying work on distributing the tasks across multiple networked devices is sound, but the context in which this research is placed doesn't make sense and is both confusing for the reader and will lead to the underlying work being misunderstood. 

I would encourage the authors to look for a more realistic use case. This is important for a journal such as Electronics, and if they wish to present their work in the context of an IoT framework. 

One example might be speech processing for smart-speaker which currently relies on sending audio data to the cloud; naturally, this has led to privacy concerns. If audio AI-processing could be distributed across multiple smart-speakers, or other networked home devices (e.g. TV, Wifi access points, etc), then smart speaker functions could be executed in a user's home ensuring privacy. Another context might be depth perception within an indoor room scene - again the local IoT computation keeps "home-data" private. 

I would also encourage more discussion in the conclusions section; there is a lengthy introductory discussion to place the work in the IoT/MEC context, but very little follow-up discussion in the conclusions section. What did we learn after implementing the MTS framework? What was the measure of success for this technique? What are the limits of scalability? (How many devices? What level of computational task can benefit?). 

In summary, the underlying work is sound, but the context and conclusions need to be significantly improved to justify publication. 

Author Response

 Response to the reviewer’s comments:

1. Does the introduction provide sufficient background and include all relevant references?

We have re-described the context specific to the environment as follows:

Advanced wireless broadband technology has introduced an unprecedented data traffic upgrade in the vehicular network (VNET). This aims to improve safety and fuel economy and reduce traffic congestion in the transportation system. To cope with these challenges, the offloading tasks to road side units (RSU) has been proposed to improve quality-of-service (QoS), although a large number of computation have been carried out during difficult deadline [1,2]. Nowadays, mobile smart devices become a common tool for social networks such as entertainment, learning, and smart life [3,4]. While mobile applications continue to emerge and computing power is becoming more and more dense, due to resource limitations of mobile devices (eg, battery life, storage capacity), the computing power of  mobile devices is still limited, making mobile users unable to achieve the same satisfaction as desktop users [5]. A more effective way to improve the performance of mobile device programs is to offload some of their tasks to the remote cloud [6,7]. However, the cloud is usually far from the mobile device, resulting in long data transmission delays between mobile devices and unpredictable results[8,9]. This is not good for mobile device programs that respond immediately. Time is of the utmost importance to mobile users, such as augmented reality apps and mobile multi-player gaming systems. To overcome the challenges mentioned above, mobile edge computing [10,11] enables mobile devices to access their internal applications and serve a variety of wireless access networks [12,13]. This approach enables computing tasks and storage capacity from the core network to be transmitted to the edge network to reduce latency. In view of some characteristics of MEC technology, multiple types of access technologies have been allowed, so vehicles can access MEC servers through various base stations (BS), such as Wi-Fi access points (Wi-Fi APs), RSU, and evolved NodeBs (eNBs).

2. Is the research design appropriate?

We have redesigned and introduced the specific model:

In this section, we discuss some models and special concept definitions. As shown Figure.1 below: A VNET consists of multiple eNBs, multiple RSUs that host the MEC server and multiple vehicles. The MEC server is placed around the edge of the core network rather than eNBS, which allows vehicles to access computing source via different wireless access methods. We suppose that the demanding vehicle is able to simultaneously offload the computing and upload its tasks to the RSU by using a full-duplex technology[17,18]. We apply a MEC-based VNET to sustain multi-vehicles. Many vehicles in the vicinity of several eNBS coverage are served by the same MEC server and expanding the range of MEC services can better meet the challenges of high-speed mobile vehicles. The eNBs that connect the entire coverage area of the MEC server are defined as the service areas of the server. In order to achieve resource utilization and ensure minimal VNET network loss, we consider a reasonable scheduling strategy to adjust computing tasks to different requesting vehicles and coordinate wireless access for vehicles through a wide range of resources. Here, multi-vehicles are treated as multi-tasking, when multiple vehicles are performing computation task offloading, there will be some impacts such as delay time and load balancing indicators on overall VNET. We need to apply the best scheduling strategy to minimize VNET loss. To avoid conflicts between multiple computing tasks when the vehicle moves, we use a multi-task scheduling (MTS) algorithm to minimize the loss of the overall network performance. Next, in the multi-task scheduling process, we model multi-task scheduling as a multi-objective optimization problem, and then try to find this Pareto optimal solution.

3.  the clarity and quality of presentation should be improved. Some abbreviations should be explained.

The specific explanation is as follows:

We first give the definition of KKT: The Karush-Kuhn-Tucker (KKT) conditions are necessary conditions that a solution to a general nonlinear programming problem must satisfy, provided that the problem constraints satisfy a regularity condition called constraint qualification. If the problem is one in which the constraint set (i.e., solution space) is convex and the maximizing (minimizing) objective function is concave (convex), the KKT conditions are sufficient. Applied to a linear-programming problem, the KKT conditions yield the complementary slackness conditions of the primal and dual problems.

4. Are the conclusions supported by the results?

The conclusions of the  mobile edge computing scheduling problem in the vehicular networks are as follows:

In this paper, we form a framework of optimal resource allocation strategy for computing in vehicular networks. We formulate the optimal computing task scheduling to minimize the VNET loss under the constraints of dynamically computational resource at RSU and the limited storage capacities as well as under the constraints of hardware deadline at end-to-end and the vehicular mobility. To avoid conflicts between tasks during vehicle mobility, we convert the multi-task scheduling problem into a multi-objective optimization problem, and then find the Pareto optimal solution. For the specific large-scale vehicular network in reality, we propose a Frank-Wolf-based MGDA optimization algorithm and extend it to the high-dimensional space. Meanwhile, we give the upper bound of the MGDA algorithm and prove that it can be computed by a backward propagation without a specific-task gradient. Finally, the experimental results show that our method is greatly improved in terms of  accuracy compared with the existing methods.

The responses are attached.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper presents an interesting approach on optimizing the multi-tasking scheduling for mobile edge-computing applications. The approach consist in identifying the scheduling problem as a multi-objective optimization problem and to find the corresponding Pareto optimal solution.

The paper is good overall. I recommend the authors to add a section where related works should be presented along with a state of the art review of this field.

The experimental part is promising, but I recommend the authors to detail and explain the results. 

Also, the clarity and quality of presentation should be improved. Some abbreviations should be explained (i.e. KKT - Karush-Kuhn-Tucker, etc.).

Author Response

Response to the reviewer’s comments: 1.Does the introduction provide sufficient background and include all relevant references?

We have re-described the context specific to the environment as follows: Advanced wireless broadband technology has introduced an unprecedented data traffic upgrade in the vehicular network (VNET). This aims to improve safety and fuel economy and reduce traffic congestion in the transportation system. To cope with these challenges, the offloading tasks to road side units (RSU) has been proposed to improve quality-of-service (QoS), although a large number of computation have been carried out during difficult deadline [1,2]. Nowadays, mobile smart devices become a common tool for social networks such as entertainment, learning, and smart life [3,4]. While mobile applications continue to emerge and computing power is becoming more and more dense, due to resource limitations of mobile devices (eg, battery life, storage capacity), the computing power of  mobile devices is still limited, making mobile users unable to achieve the same satisfaction as desktop users [5]. A more effective way to improve the performance of mobile device programs is to offload some of their tasks to the remote cloud [6,7]. However, the cloud is usually far from the mobile device, resulting in long data transmission delays between mobile devices and unpredictable results[8,9]. This is not good for mobile device programs that respond immediately. Time is of the utmost importance to mobile users, such as augmented reality apps and mobile multi-player gaming systems. To overcome the challenges mentioned above, mobile edge computing [10,11] enables mobile devices to access their internal applications and serve a variety of wireless access networks [12,13]. This approach enables computing tasks and storage capacity from the core network to be transmitted to the edge network to reduce latency. In view of some characteristics of MEC technology, multiple types of access technologies have been allowed, so vehicles can access MEC servers through various base stations (BS), such as Wi-Fi access points (Wi-Fi APs), RSU, and evolved NodeBs (eNBs).

2.Is the research design appropriate?

We have redesigned and introduced the specific model: In this section, we discuss some models and special concept definitions. As shown Figure.1 below: A VNET consists of multiple eNBs, multiple RSUs that host the MEC server and multiple vehicles. The MEC server is placed around the edge of the core network rather than eNBS, which allows vehicles to access computing source via different wireless access methods. We suppose that the demanding vehicle is able to simultaneously offload the computing and upload its tasks to the RSU by using a full-duplex technology[17,18]. We apply a MEC-based VNET to sustain multi-vehicles. Many vehicles in the vicinity of several eNBS coverage are served by the same MEC server and expanding the range of MEC services can better meet the challenges of high-speed mobile vehicles. The eNBs that connect the entire coverage area of the MEC server are defined as the service areas of the server. In order to achieve resource utilization and ensure minimal VNET network loss, we consider a reasonable scheduling strategy to adjust computing tasks to different requesting vehicles and coordinate wireless access for vehicles through a wide range of resources. Here, multi-vehicles are treated as multi-tasking, when multiple vehicles are performing computation task offloading, there will be some impacts such as delay time and load balancing indicators on overall VNET. We need to apply the best scheduling strategy to minimize VNET loss. To avoid conflicts between multiple computing tasks when the vehicle moves, we use a multi-task scheduling (MTS) algorithm to minimize the loss of the overall network performance. Next, in the multi-task scheduling process, we model multi-task scheduling as a multi-objective optimization problem, and then try to find this Pareto optimal solution.

3. the clarity and quality of presentation should be improved. Some abbreviations should be explained.

The specific explanation is as follows: We first give the definition of KKT: The Karush-Kuhn-Tucker (KKT) conditions are necessary conditions that a solution to a general nonlinear programming problem must satisfy, provided that the problem constraints satisfy a regularity condition called constraint qualification. If the problem is one in which the constraint set (i.e., solution space) is convex and the maximizing (minimizing) objective function is concave (convex), the KKT conditions are sufficient. Applied to a linear-programming problem, the KKT conditions yield the complementary slackness conditions of the primal and dual problems.

4.Are the conclusions supported by the results?

The conclusions of the  mobile edge computing scheduling problem in the vehicular networks are as follows: In this paper, we form a framework of optimal resource allocation strategy for computing in vehicular networks. We formulate the optimal computing task scheduling to minimize the VNET loss under the constraints of dynamically computational resource at RSU and the limited storage capacities as well as under the constraints of hardware deadline at end-to-end and the vehicular mobility. To avoid conflicts between tasks during vehicle mobility, we convert the multi-task scheduling problem into a multi-objective optimization problem, and then find the Pareto optimal solution. For the specific large-scale vehicular network in reality, we propose a Frank-Wolf-based MGDA optimization algorithm and extend it to the high-dimensional space. Meanwhile, we give the upper bound of the MGDA algorithm and prove that it can be computed by a backward propagation without a specific-task gradient. Finally, the experimental results show that our method is greatly improved in terms of  accuracy compared with the existing methods.

The responses are attached.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

This work now has a good use-case that makes sense and thus is much closer to being publishable. 

The conclusions section is still weak. The authors describe a "large scale vehicular network" and state that "Many vehicles in the vicinity of several eNBS coverage are served by the same MEC server, and expanding the range of MEC services can meet the challenges of high-speed mobile vehicles better";  but there is no discussion on what this means, or how their solution might scale to 100s or 1000s or vehicles with multiple servers. It seems that the use case investigated is a limited one, rather than a fully multi-vehicle, multi-task, multi-server system. This would be fine for preliminary investigation of this topic, but the authors need to clarify and discuss the limitations of this study and, ideally, propose how this work could be extended to a more practical VNET scenario.  

Author Response

We updated the manuscript by adding a future works, as follows:

In this paper, we study how to implement an effective and reasonable scheduling strategy in the vehicular network to minimize the performance loss of the entire network. The significance of its research is that classification-based tasks have been well promoted in the field of deep learning models, but there are certain limitations. For example, as the size of the network increases and the number of vehicles increases, there will be phenomena such as traffic congestion and insufficient cache. How to solve these problems in the super network is considered in the future. At the same time, in the vehicular network, when the vehicle moves on the roadside base unit, privacy information will be revealed. How to design the encryption scheme is also considered in the future.