Task Offloading of Deep Learning Services for Autonomous Driving in Mobile Edge Computing

Round 1

Reviewer 1 Report

The paper targets autonomous driving applications and proposes a task offloading algorithm for deep learning services between mobile vehicles and edge computing servers. The paper considers several representative tasks deployed in autonomous driving and use a real-time server monitoring system and scheduler to efficiently utilize distributed computing resources to balance system loads and requests processing delay. The offloading problem is formalized as a optimization problem and eveluations demonstrate the performance of the proposed design.

The paper targets a timely topic and meaning problem. The problems are well formalized. Also, the design is comprehensive. Both tasks characteristics, hardware resources and algorithms optimizations are considered. However, there are several comments, and it is advised to address them.

First, regarding the scheduler, what kinds of specific scheduling algorithms are used? The paper simply uses an equation to describe the tasks arrival/departure, and queue backlogs. It is still unclear how the queue/scheduler works. How does it push/pop task and what principles does it use to schedule tasks?

Second, how do the hardware resources instantiated in the offloading/optimal formalization?

Third, it is advised to discuss 10.1109/RTAS54340.2022.00031, which demonstrates the tasks offloading and cooperation between vehicle and edge-based infrastructure.

Forth, regarding the different tasks, they are typically heterogeneous, i.e., they are using different CPU/GPU/memory, and may be computing-/memory-intensive. How will this impact the offloading algorithm?

Looks good to me

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This report highlights a paper on the utilization of mobile edge computing and task offloading for autonomous driving, and proposes a novel task offloading algorithm based on Lyapunov optimization. The paper addresses the limitations of existing methods by incorporating real-time monitoring, analyzing computational complexity, and employing Lyapunov and Lagrange optimization to strike a balance between system stability and user requirements.

The experimental results demonstrate the effectiveness of the proposed algorithm. The system queue backlog remains stable, and the average processing times for object detection, driver profiling, and image recognition are well within acceptable limits. With processing times of 0.4231 sec, 0.7095 sec, and 0.9017 sec, respectively, the algorithm ensures that deep learning applications can meet critical deadlines.

By efficiently utilizing distributed computing resources, the algorithm enables autonomous driving systems to handle complex and heavy applications. Its ability to adapt to dynamically changing environments sets it apart from previous methods. This research represents a significant advancement in the field and paves the way for more efficient and stable autonomous driving systems in the future.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper has a very good level. I suggest explaining a little more about the study of real-time analysis, specifically on response times and maximum deadlines.

did you use containers in rt-linux?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The main question addressed by the research was how to develop a task offloading algorithm for deep learning services in an autonomous driving environment that ensures system stability and minimizes task processing delay.

This topic is not original as there are already various perspectives on the task offloading policies, including ones based on the AI, but these strategies require a lot of time and training data necessary for supervised learning. The authors addressed this specific gap in the field by formulating task offloading steps using the Lyapunov optimization and Lagrange optimization framework.

Compared with other published material this approach enables decision-making with minimal observational information, such as queue backlog, arrival data size, and server computing resources capability, without requiring knowledge of the system's probability distribution. The mathematical formulation and proof of optimization and task offloading steps ensure performance. Consequently, the researchers employ Lyapunov optimization to balance the trade-off between profit and cost in task offloading.

 

Before this paper is ready for publication, the authors need to expand the description of Driver profiling, image recognition and object detection. 

Regarding driver profiling, what is its purpose in the autonomous driving application? Driver profiling is used by HR managers of logistics and transportation companies to determine if they should hire a candidate or not based on his personality (e.g. aggressive, normal, or drowsy driver), thrill seeking or anxiety, law conformity or speed limit violations. If a vehicle drive itself, who is being profiled, and why? perhaps drivers of surrounding vehicles, but how would that influence the behavior of a self-driving car?

Second, they need to clarify the difference between image recognition and object detection, and why these two actions must be separate.

 

In their future work, the authors should try to test the proposed methodology in realistic conditions, in this paper they used CARLA simulator and the AWS deep learning solution, but what was the connection between these two? I assume a high bandwidth optical internet. But how would a real autonomous vehicle access AWS server, using 5G mobile network perhaps??? Data transmission delay (As pointed out in the related work line 120) should be further investigated.  

 

Conclusions are consistent with the evidence and arguments presented in the paper and address the main question of task offloading feasibility. However, the conclusion section should be expanded with more detailed reflections on the methodology used in the paper.

 

References appropriately describe the state of the art in the autonomous driving applications and current challenges.  

 

Tables in this paper are well placed and very effectively summarize key data.

Figure 1. is OK.

The rest of the figures need to be enlarged in order to be readable (font size on figures must match the font size in the rest of the paper)

For the reference, authors should print their article on a standard A4 paper and see if they can distinguish letters (especially in Figure 7)

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The revised paper is significantly improved. The responses are detailed and solid, and addressed my concerns. In particular, the offloading and scheduling algorithms are discussed in detail, which further highlights the contributions of the paper. Also, the response to point 4 makes sense and is convincing. The task and each node are heterogeneous, but all nodes are homogeneous, and the tasks are scheduled among different nodes. Overall, I recommend this paper to be published.