CCoW: Optimizing Copy-on-Write Considering the Spatial Locality in Workloads

Round 1

Reviewer 1 Report

MDPI Electronics

The following is an overview of the article CCoW: Optimizing Copy-on-Write Considering the Spatial Locality in Workloads (electronics-1545825). In this study, author(s) proposed the Coverage Copy-on-Write (CCoW), a locality-based Copy-on-Write techinique using the coverage to solve the problem. CCoW could reduce the number of the write faults and improve the performance by identifying the pages where the write faults are occured frequently in the process, and pre-copying the entire pages in a single write fault.

However, some points must be highlighted so that the author(s) can review and submit in another round of review: The following corrections are considered to be beneficial for the strengthening of the article.

1. The Conclusions should be reviewed again. The original aspect of the study and its difference from other studies should be clearly explained. (The conclusion should be explored better and it needs to contemplate the eventual restrictions of the developed technique to address future works in this area.)
2. The abstract must be make strong. Abstract should be reviewed again.
3. Some sentences have spelling errors. (Punctuation marks, spaces, etc.). Some places should be left space. Please check all sentences in article.
4. It has been a comprehensive study in the literature in recent years. If there are more current literature studies, these should be examined in detail and added to the literature section.
5. The authors should compare the results of their method with those of previous studies. As mentioned in the literature, there are several methods with very high accuracy, even better than the proposed method. Author(s) can do compare table (A new table can add about previous studies to result section.). This subject is very important.
6. The motivations of the proposed method are not clear. Which problem does the proposed method attempt to solve? Why the other existing diagnosis methods failed to solve it? What are the advantages of the proposed method compared to other methods? Those should be illustrated more clearly.
7. Carefully check all grammatical error. Still, the English language should be improved. I suggest asking for help from a native English.

I think it is ACCEPTABLE after the MAJOR Revisions mentioned.

Comments for author File: Comments.pdf

Author Response

We appreciate the reviewer for the invaluable comments. We incorporated the comments and improved the manuscript greatly according to the suggestion.   Abstract and Conclusion (Coment 1 and 2): Thanks for the comments. We frankly admit that the sections were not properly developed in the previous submission. We completely rewrote the sections to better introduce and conclude our work, as the reviewer pointed out.   Editorial issues (Comment 3 and 7): Thanks for the comments. We revised the entire manuscript through an editorial service to correct many grammatical errors, misspells, typos, and punctuations.   Comparison to other work (Comment 4, 5, and 6): We appreciate the reviewer for the suggestion. We totally agree with the reviewer's opinion suggesting the performance comparison. To the authors' best knowledge, however, there is no study that is directly comparable to the proposed CCoW. The work discussed in Section 2 is not actually comparable to CCoW. Specifically, Baumann et al. [20] proposes to replace fork system call with a new system call simplifying process creation, but applications cannot make the clone of memory data using their proposed system call, making it infeasible to be compared with CCoW. Zhao et al.[19] proposes to allow copy-on-write to the `page table', and in that sense, this scheme is orthogonal to the proposed CCoW which discusses on the copy-on-write to the `data' pages. As a matter of fact, CCoW can be applied to the scheme that Zhao et al. suggests, and the scheme can be applied to the system with CCoW, vice versa. In addition, huge page-related work have been discussed. We introduced those literatures in recent years on lines 111--126 on page 3. They allow to identify high-locality memory regions with high accuracy as the reviewer pointed out. However, copy-on-write is processed in a per base page manner. Thus, on the write-intensive workloads where CoW bounds the application performance, they perform just similar to the baseline configuration with CoW. We mentioned these on lines 225--233 on page 5--6 to differentiate our work to them. To support this idea and claim, we measured the performance of the system with the transparent huge page support, and present the result on Figure 5 and lines 430-435 on page 12 (with microbenchmark), and Figure 7 and lines 486--491 on page 15 (with Redis). Specifically, THP allows some performance improvement, however, the improvement is limited under the write-intensive workload as huge pages are likely to be split into base pages for CoW. As a result, we can verify that CCoW outperforms the THP-enabled configuration.

Reviewer 2 Report

The authors suggest Coverage Copy-on-Write (CCoW) to improve performance. The proposed model was implemented on the basis of Linux. However, the authors should pay attention to the following notes: 1. Abstract: It is necessary to justify that the subject of the work is suitable for the Electronics magazine. 2. Review part: In addition to virtualization systems, it should also be noted specialized software solutions for optimizing the operation of algorithms (doi: 10.1109 / SYNCHROINFO49631.2020.9166067) and other optimization methods (10.26599 / TST.2019.9010034.) And one of the leading articles on memory capabilities in relation to neural networks ( http://proceedings.mlr.press/v48/santoro16.pdf) 3. Lines 2, 4: What do the extra parentheses after the word "fork" mean? Can you explain please. 4. Line 5: We proposed ... We should try not to use personal pronouns in our work. It was proposed or text proposed ... 5. Line 37: What is TLB? 6. Figures 3-5, 7: Explain why Normalized Performance can reach values greater than 1? 7. Page 14: Table 1 - The table heading must be before the table, not after.

Author Response

We appreciate the reviewer for the invaluable comments. We incorporated the comments and improve the manuscript as suggested.   Respond to Comment 1, 3, 4, 5, and 7: We revised the entire manuscript through an editorial service to correct many grammatical errors, misspells, typos, and punctuations. Also, "Abstract" is completely rewritten to better introduce our work. Specifically, some unclear terms (e.g., fork and TLB) have been removed from the abstract and discussed later in details in Section 2.1 on page 3. We had put parentheses after the word "fork" to explicitly indicate the "fork system call", but we changed the style of presentation to remove the unclearness that the reviewer pointed out.   Respond to Comment 2: Thanks for introducing the interesting work. For sure, copy-on-write has been discussed in the context of virtualization, for example, to duplicate a VM instance more efficiently or to optimize VM migration across network. However, we are afraid that our proposed scheme has little relevance to some of the papers that the reviewer mentioned. Specifically, our work is focusing on improving copy-on-write in write-intensive workloads. The mentioned papers are, however, discussing on accelerating neural network training and inference, where copy-on-write does not incur. It would be really appreciated if the reviewer confirms one's concern. For paper 10.26599 / TST.2019.9010034, it discusses to improve the copy-on-write from the docker container. The proposed scheme is, however, only applicable to the `file-backed pages' storing the contents from files, whereas CCoW is focusing on `anonymous pages' storing data for the heap and stack. We explain their difference in Section 2.1 on page 3--4.   Respond to Comment 6: We apologize the confusing. The performance is normalized to the performance of `CoW-only' configuration. In the figures of the previous submission, the values for the baseline configuration were the last ones, not the first one, among the bars, and it may mislead the reviewer. We plot the figures again so that the baseline comes first followed by others. Also we meant the normalized performance to the performance (i.e., operations per second) of the configuration divided by the performance of the baseline configuration, so the higher the value is, the higher performance the system exhibits.

Round 2

Reviewer 1 Report

Dear Editor;

The author(s) made all the corrections mentioned (electronics-1545825_R1 - CCoW: Optimizing Copy-on-Write Considering the Spatial Locality in Workloads).

The length of the paper is enough in terms of a scientific paper.