Analytical Energy Model Parametrized by Workload, Clock Frequency and Number of Active Cores for Share-Memory High-Performance Computing Applications
, Carlos Valderrama 1, Pierre Manneback 1 and Samuel Xavier-de-Souza 2
Round 1
Reviewer 1 Report
The paper proposes an energy model based on operating frequency and number of cores for a shared memory system.
In my opinion, the article is written in a very clean and pleasant way. Reading this article was pleasant and gradually introduced the reader to the issues presented. The objectives as well as the research methodology are presented in a clear and transparent manner.
Nevertheless, I have a few small comments that are merely optional and do not detract from the quality of the article:
- Equation 11 and 12 - in order to show the relationship between e.g. P and f, a symbol very similar to the symbol α was used. Previously, the symbol α was used in other equations. As a result, the reader may mistake this notation. I suggest you use a different math symbol.
- Line 271 - space after "time.Figure" is missing. It should be "time. Figure".
- Table 1 and other - shows numerous measurements for various applications. It is worth indicating the source where you can find the descriptions of these applications, and preferably their source codes, in order to avoid mistakes.
- Abbreviations - it is worth adding dashes, e.g. "MPE - Mean Percentage Error".
Author Response
Please see the attachment with our responses!
Thank you!
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The paper presents a model to predict the energy consumption by parameters of frequency, number of cores and workload etc. The prediction can be useful for energy saving purposes which is interesting.
However, the writing needs to be improve significantly before submitting for publication. For example, the definition of DVFS in the introduction and also in the abbreviations is not correct.
Author Response
Please see the attachment with our responses!
Thank you!
Author Response File: Author Response.pdf
Reviewer 3 Report
The work presented in your paper proposes, describes, and validates an analytical energy model for share-memory high-performance computing applications, obtaining interesting results when compared with well stablished methods like DVFS and DPM. The subject is timely, as it concerns the energy and power reduction and optimization, the contents of the paper is comprehensive, and the set of obtained data tests is exhaustive. However, I have some comments and observations to take into consideration before the final acceptance of your article.
- Introduction
Line 42: References [12-14] are quite old, it would be desirable to add more recent ones.
Line 61: A reference should be added to illustrate the introduction of DVFS algorithms in 1994.
Lines 71-79: Before the paragraph which starts in line 80, it is necessary to detail the main objectives and contributions of your paper. Even if you already present some ideas in these lines, in my opinion, the interest of your work is not clearly highlighted. Perhaps, a list summarizing these objectives could be helpful.
- Theoretical background
My major remark regarding this section is the place where it appears in the paper. Indeed, I think that this section should appear after the current Section 3, just before Section 4. So, the new suggested structure of the paper would be: 1. Introduction; 2. Related work; 3. Theoretical background; 4. Modeling energy with performance and power; 5. Experimental validation and 6. Conclusion. In this way, you first conduct a state of the art of the existing methods, next you present the theoretical background to finally introduce your method, which considers power, performance, and energy concerns.
- Related work
In this section, you could add the end a table summarizing the most relevant differences between your method and the presented works, to point up the improvements that you propose.
- Modeling energy with performance and power
Section 4.1: Could you quantify the precision and the error of the presented model, which is based on the power consumption of the transistors, when compared with real systems? This is a very interesting approximation whose limitations merit to be explained.
- Experimental validation
Lines 223-229: When you say that “SVR was chosen as the most representative because it performed best in our tests without aggressive fine-tuning.”, you should also show the experimental results that support this decision, including for example a table with the obtained results of the comparison of your method with the enumerated machine learning approaches. Alternatively, a reference citing your previous work can be also provided.
Line 224: Why the “Userspace governor” has been finally employed?
Line 273: When you say “Figure 4 shows that the assumption was reasonable”, what does “reasonable” mean for you? Could you quantify this?
Lines 273-276: Here, I think that a table showing all the obtained results is mandatory, including error quantifications.
Table 1: The information included in this table is not very clear. How can we see the variation of the number of instructions when changing the number of cores? Unless I am wrong, the number of cores is not visible in the table. I have the same remark concerning Table 2.
Table 3 is never presented in the paper.
Figures 6, 7 and 8: Trained values are not very visible in the graphs. How can we observe the “minimum energy” value in all graphs?
- Conclusion
Could you please present and explain the main limitations of your study? In which measure could the obtained results be generalized to other node architectures and configurations, different from that employed to carry out your experimental tests?
References
Please verify the style and format of references since there are several mistakes.
Author Response
Please see the attachment with our responses!
Thank you!
Author Response File: Author Response.pdf
Reviewer 4 Report
The topic of the paper is relevant and interesting. The authors proposed an energy model based on the operating frequency and the
number of cores for a shared memory system. The paper is well written and organized. All methods and approaches, which were used were explained and described. I agree with the authors, that optimization energy consumption increases efficiency usage of computers in total. The authors used a lot of methods and analyses observed parameters. But my opinion will be useful to develop a simplified algorithm for choosing optimal value configuration computer and organizing his optimal operational mode. Maybe adding some flowcharts gives the possibility to summarise obtained results and formulate a new methodology.
Author Response
Please see the attachment with our responses!
Thank you!
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Thank you for your corrections and clarifications following my comments. The current paper has been substantially improved compared with your previous version, and the main issues have been well addressed.
