Fuzzy-Based Thermal Management Scheme for 3D Chip Multicores with Stacked Caches

Round 1

Reviewer 1 Report

This paper describes a novel thermal management scheme based on fuzzy logic.

The paper is well structured and the conclusions are supported by the experimental results.

I have only two main points that should be addressed before paper publication:

Algorithm 1 should be better explained in order to allow the results reproduction. In particular, I suggest to write an extensive explanation which makes reference to the corresponding line in algorithm 1.   On line 218 Authors make reference to Figure 5as the “defuzzification membership function”. However, this figure is already referenced as “Membership function of the logistic variables”. Can the author better explain this point?

Please, carefully revise the English style since there are different sentences which are hard to understand. There are also typos (“Therefor” instead of “Therefore”) and verbs in the short form that should be avoided (“Hasn’t”).

Author Response

Dear Reviewer

    Thanks for the insightful suggestions and comments work on our manuscript. We have studied the comments carefully and made corrections accordingly. Point by point responses to your comments is listed below.

 

Point 1: Algorithm 1 should be better explained in order to allow the results reproduction. In particular, I suggest to write an extensive explanation which makes reference to the corresponding line in algorithm 1.

Response：Thanks for the comment. We are sorry that we have not illustrated Algorithm 1 clearly. Therefore, we add detail explanation for Algorithm 1 in Section 3.1.

 

Point 2: On line 218 Authors make reference to Figure 5as the“defuzzification membership function”. However, this figure is already referenced as “Membership function of the logistic variables”. Can the author better explain this point?

Response：Thanks for the comment. We are sorry that we made a mistake on line 218 in previous version. We corrected the error in the resized version. The membership functions of the input and the output are shown in Fig. 5. During fuzzification stage, the values of e(n) andΔT(n) will be mapped into their membership functions for the value applied to certain fuzzy set will be calculated. The fuzzy inference stage makes the decision for the output based on a set of fuzzy rules. For the defuzzification  module, we also adopt the mean-of-maxima method. Finally, the fuzzy output will be combined together with mean-of-maxima method and converted into a frequency ratio of the processor.

 

    We have tried our best to improve the manuscript and made some changes in the manuscript. On the whole, these changes will not influence the content and framework of the paper.

    We really appreciate for your work on our manuscript, and hope that the correction will meet with approval.

 

    Best regards!

 

                                                                          The authors

                                                                 February 13th, 2020

 

Reviewer 2 Report

This is an interesting research paper. There are some suggestions for revision.

1. The motivation is not clear. Please discuss how cache and processor power consumption affect the thermal control respectively.

2. Please compare the pros and cons of existing solutions.

3. Please highlight your contributions in introduction.

4. In line 83-84, it mentions "We assume that the core layer closeted to the heat sink consists of four processor cores and every cache layer has 16 cache bank". If the architecture is not same as the assumption, can it get the same result?

5. Please specify the input and output of Algorithm 1.

6. The proposed solution has several parameters. Please show how to set/get these parameters.

7. For input and output of section 3.2, what is your contribution in Eq. 4-9.

8. For the fuzzy membership functions shown in Fig. 5, where are the values from or how to get the values?

9. Please explain how to get Tab. 2.

10. In experiment, please discuss the relationship among baseline, FLC, and FBTM.

11. The experiment results are not convincing. Please compare the proposed solution with existing solutions.

Author Response

Dear Reviewer

    Thanks for the insightful suggestions and comments work on our manuscript. We have studied the comments carefully and made corrections accordingly. Point by point responses to your comments is listed below.

 

Point 1: The motivation is not clear. Please discuss how cache and processor power consumption affect the thermal control respectively.

Point 2: Please compare the pros and cons of existing solutions.

Point 3: Please highlight your contributions in introduction.

Response：We are sorry that the description of the motivation and our contribution is not clear. Therefore, we reorganized the Introduction to discuss the affection of the cache and processor power consumption and our contributions, which were highlighted in the paper.

In the Introduction, we compare the pros and cons of existing solutions and our prior work, then propose our thermal management scheme.

 

Point 4: In line 83-84, it mentions "We assume that the core layer closeted to the heat sink consists of four processor cores and every cache layer has 16 cache bank". If the architecture is not same as the assumption, can it get the same result?

Response：For simplicity of the problem, we target an example architecture. The proposed method can be easily extended to multi-core processor architecture.

 

Point 5: Please specify the input and output of Algorithm 1.

Point 6: The proposed solution has several parameters. Please show how to set/get these parameters.

Response：The input of Algorithm 1 is the application workload and the output of the Algorithm 1 is the reconfigured cache. We are sorry that we have not illustrated Algorithm 1 clearly. Therefore, we add detail explanation for Algorithm 1 in Section 3.1.

 

Point 7: For input and output of section 3.2, what is your contribution in Eq. 4-9.

Response：The goal of the design is to limit the temperature near the threshold temperature. From Eq.4-9, we can confirm the input parameters includes: 1) the difference between the measured temperature and the temperature threshold and 2) the temperature changing rate.

 

Point 8: For the fuzzy membership functions shown in Fig. 5, where are the values from or how to get the values?

Response：We are sorry that we have not illustrated the fuzzy membership functions clearly. We get the parameters by simulation using HotSpot. We add detail explanation for the fuzzy membership functions in Section 3.2.

 

Point 9: Please explain how to get Tab. 2.

Response：Table 2 shows the fuzzy rules, which are designed by the commonsense of people. Low thermal error means that the temperature is relatively near the threshold temperature. Then positive and zero thermal change rates may bring the temperature to exceed the threshold temperature. The high-frequency level may be set in the next time period. A high thermal error means that the temperature is below the threshold temperature. Then negative and zero thermal change rates mean that the temperature is barely increasing. The low-frequency level may be set in the next time period. When the thermal error is normal, a positive thermal change rate may bring a high-frequency level, while a negative thermal change rate may bring a low-frequency level.

 

Point 10: In experiment, please discuss the relationship among baseline, FLC, and FBTM.

Response：We discuss the relationship among baseline, FLC and FBTM in Section 4. The baseline is a base method adopting DVFS technique. FLC is a method using the fuzzy-based control policy without a dynamic cache reconfiguration module and FBTM is our proposed thermal management method using a fuzzy-based control policy and dynamic cache reconfiguration module.

 

Point 11: The experiment results are not convincing. Please compare the proposed solution with existing solutions.

Response：DVFS has been a crucial technique in utilizing the hardware properties of processors to drop energy dissipation. Our contribution of the paper is that we propose a fuzzy-based control policy which is based on fuzzy logic. Besides that, we consider the impact of cores and on-chip caches in parallel. Therefore, in experiments we compare the proposed method with an existing DVFS technique and the method only adopting fuzzy-based control policy. The experimental results show that the proposed method achieves 3 degrees reduction on average in temperature and gives a 41% reduction on leakage energy compared with the existing DVFS method. The detail experiment results are discussed in Section 4.

 

   We have tried our best to improve the manuscript and made some changes in the manuscript. On the whole, these changes will not influence the content and framework of the paper.

   We really appreciate for your work on our manuscript, and hope that the correction will meet with approval.

    Best regards!

 

                                                                                    The authors

                                                                               February 13th, 2020

 

Round 2

Reviewer 2 Report

All my concerns have been addressed. This paper is ready for publication.