Partitioning DNNs for Optimizing Distributed Inference Performance on Cooperative Edge Devices: A Genetic Algorithm Approach
Round 1
Reviewer 1 Report
Review on manuscript entitled “Partitioning DNN for Optimizing Distributed Inference Performance on Cooperative Edge Devices: A Genetic Algorithm Approach”.
Presented work is interesting and topic is emerging with many researchers working in the field. However, presented test case is not well documented and needs improvements before publication.
1) Problem formulation must be exact with used problem parameter variations, objectives and constrains. Probably table with actual numerical values will be useful.
2) Algorithm implementation must be defined with actual numbers used. All algorithm parameters e.g. number of generations, chromosome size, loop repeats, must be declared.
3) GA is stochastic method. How did you compare execution times which probably correspond to different objective function values?
Assuming 1-3, presented research must be reproducibly documented. An independent researcher must be able to repeat your results, so all relevant data must be presented in your work. Alternatively, source code could be provided in repository as github.
Author Response
Dear reviewer,
Thanks very much for taking the time to review this manuscript. I really appreciate all your comments and suggestions! According to your suggestions, I have reorganized the description of the experiment setting in section 5.1 and added one new paragraph (line 305 to 313) to show all the parameter values set in our experiments, including population size, the number of generations, stop conditions, etc. Please find my revisions in the re-submitted files.
Besides, we carried out many experiments according to the scale of the problem and finally set the iteration times of the genetic algorithm to 200 in our experiment.
Thank you again for your valuable comments.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors could work still on a mathematical model to reduce working on CPU rather than GPU.
The authors could have worked on 3D images which could have made a good research paper.
Author Response
Dear reviewer,
Thanks very much for taking the time to review this manuscript. I really appreciate all your suggestions! This paper mainly takes DNN model segmentation as an example. It focuses on generating a more effective collaborative-inferencing scheme in a distributed environment, which can fully use the decentralized computing resources in the edge environment to improve the efficiency and effect of edge intelligent applications. As you mentioned, using mathematical methods to reduce the work on the CPU is also of great research significance. We will take further study on it.
In addition, the recognition of 3D images is more challenging compared with the current experimental scenarios. We will also consider improving our method in 3D image recognition in the future.
All your valuable suggestions are listed in our further research plan, and we added relevant descriptions at the end of the conclusion part (line 401-403). Please find my revisions in the re-submitted files.
Thank you again for your valuable comments.
Author Response File: Author Response.pdf
Reviewer 3 Report
The topic of the article is really interesting, but the following issues have to be adressed and solved:
- in the abbreviation there is a part in 8th line "our algorithm can result in about 1 3x shorter running time than the basic GA for getting a better deployment." - do you mean 13x time or 1.3x time ?
- this could be profitable to add in the introduction section the following sources: https://www.mdpi.com/1424-8220/22/7/2496 and https://www.mdpi.com/2071-1050/14/6/3312
- the same issue as for the first indicated point is abuout the line 47,
- Chapter number 2 should be rather named as Literature review, no Related work,
- layer L5, 5 should be indexed, - line 174,
- Such deployments will lead to extra network bandwidth and equipment energy consumption caused by repeated transmission between devices - line 176, 177 - could you be more precise and give some approx percantage values,
- line 182, please dont use personal form such as We, please use eg. The work proposes ...
- above issues is about the whole manuscript, please correct this, eg. line 190, 196..
- please improve the visibility of text on Figures 4 and 5 by increasing the font size,
- in the conculsion part, maybe it should be worth also to mention comparison and the results abut device average energy consumption.
Author Response
Dear reviewer,
Thanks very much for taking the time to review this manuscript. I really appreciate all your comments and suggestions! Please find my itemized responses below and my revisions in the re-submitted files.
In line 9 and line 52, I changed the abbreviation "1 3x“ to "1 to 3 times".
In the first line of the introduction (lien 13 - 15), I added the recommended references, https://www.mdpi.com/1424-8220/22/7/2496 and https://www.mdpi.com/2071-1050/14/6/3312.
The name of section 2 is changed to “Literature Review”.(line 65)
In addition to the subscript problem in original line 147, I re-checked similar problems throughout the paper and modified them accordingly.
To make the description more clear in line 178 and 179 (original line 176 and 177), I added an example to explain. (line 188-191)
As you mentioned, to avoid using personal forms such as we, I revised all such descriptions, including we and our.
I also increased the font size in both Fig. 4 and Fig. 5 to improve text visibility.
At last, I have added the description about comparison on device average energy consumption in line 397-400.
Thank you again for your valuable comments.
Author Response File: Author Response.pdf
Reviewer 4 Report
The research proposes a Genetic Algorithm based dynamic DNN partitioning and deployment system model to represent the actual application requirements of distributed DNN inferencing in an edge environment.
The research has an average novelty with significant content targeting the most researched area currently of machine learning. The research is scientifically sound which interest the reader. The methodology is understandable and well presented and interesting.
The only challenge I was having is why the exhaustive search algorithm is said to be optimal? Also it will be good to label the exhaust search algorithm as Optimal in the graphs, because future researchers when reading this article, might assume the exhaustive search is not an optimal algorithm but Just another compared algorithm that performs better than your algorithm. Also, put legends in all the graphs such that each graph has a legend.
Author Response
Dear reviewer,
Thank you very much for your approval of our work.
As you mentioned, this paper uses the exhaustive method to find all possible solutions to select the absolute best solution as the baseline to compare. However, the description was not clear enough, which resulted in certain ambiguity. I've replaced all "Exhaustive Method" with "Optimal Value" in Fig. 4 and Fig. 5 and modified the corresponding comparison description (line 342-350). Also, legends have been added to all figures. Please find all my revisions in the resubmitted files marked up by the "Track changes" function.
Thank you again for your valuable comments.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
I am satisfied with the new proofread revision, good work has been done.
Authors may add some more details on actual problem formulation and one optimal result for clarity.
Author Response
Dear reviewer,
Thank you very much for your comments and suggestions.
According to your suggestions, I have revised the first two paragraphs in Section 3, i.e., System Model and Problem Formulation. To make the problem formulation clearer, I re-descript the motivation and system model with video-based fall detection as an example. Please find all my revision details in the attached PDF in track-changing mode.
Thank you again and best regards.
Author Response File: Author Response.pdf
Reviewer 3 Report
I have no further comments for the authors. The work is good enough. Thank you.
Author Response
Thank you very much.
