36 lines
5.0 KiB
Markdown
36 lines
5.0 KiB
Markdown
type:: [[REVIEWS]]
|
||
tags::
|
||
year:: 2024
|
||
venue:: [[TOSEM]]
|
||
full-title:: Enhancing Energy-Awareness in Deep Learning through Fine-Grained Energy Measurement
|
||
date-start:: [[11-01-2024]]
|
||
date-submitted:: [[16-03-2024]]
|
||
external-links::
|
||
status:: [[DONE]]
|
||
deadline-submission:: [[29-02-2024]]
|
||
file:: [02-29-TOSEM-2023-0413_Proof_hi.pdf](zotero://select/library/items/IDLWYMWL) {{zotero-imported-file IDLWYMWL, "02-29-TOSEM-2023-0413_Proof_hi.pdf"}}
|
||
todoist:: https://app.todoist.com/showTask?id=7442180975
|
||
|
||
- [[Highlights]]
|
||
- [[Comments]][]()
|
||
- SUMMARY: The paper discusses energy consumption measurement in deep learning frameworks, with a focus on fine-grained measurement at the API level. The paper introduces FECoM (Fine-grained Energy Consumption Meter) to profile DL APIs from an energy perspective. FECoM consists of two main components for static instrumentation of the input program to measure and for the energy consumption of the isolated APIs that are of interest to the measurement. The paper demonstrates FECoM's capability for fine-grained energy measurement through its assessment on TensorFlow.
|
||
|
||
COMMENTS: The paper is relevant and well-written, but there's room for improvement in its presentation. Here are my suggestions:
|
||
- The Related Work section should be moved earlier in the paper to provide motivation for the work after presenting the state of the art. Table 2 is a good way to address the limitations of existing techniques and motivate the use of FECoM. I suggest moving Table 2 before the approach section and adding descriptive text to introduce and motivate the comparison features. Additionally, technologies like Monsoon should be included in the table to present a complete view of the field.
|
||
- By taking into account the done for RQ3 and Table 2, there is room for defining a set of criteria that users can consider when selecting the energy measurement technique that best fits their particular needs. For instance, if the expected granularity is at the system level, FECoM might not be necessary, or if the acceptable sampling rate is a matter of seconds, CodeCarbon might be a better choice.
|
||
- Regarding RQ1, the summary is good, but it would be helpful to explicitly discuss the practical implications of the results and how they can be leveraged.
|
||
- The description of Equation (1) needs improvement. Providing an example related to Listing 1 to show the different components of the equation would make it easier to understand.
|
||
- The negative value for RAM energy consumption in some projects is unclear. Please clarify.
|
||
-
|
||
- Original
|
||
- SUMMARY: The paper is on the energy consumption measurement in deep learning frameworks, particularly focusing on fine-grained measurement at the API level. In particular, the paper introduces FECoM (Fine-grained Energy Consumption Meter) to profile DL APIs from energy perspective. The architecture of FECoM consists of two main components for the static instrumentation of the input program to measure, and for the measurement of the energy consumption of the isolated APIs that are of interest for the measurement. FECoM has been assessed for TensorFlow, and its capability for fine-grained energy measurement has been demonstrated.
|
||
- COMMENTS: The paper is on a relevant topic. Overall the paper is well written even though I have some recommendations to improve the presentation of the paper:
|
||
- My main comment is related to the Related work section. In particular, I would move this section earlier in the paper to motivate the work after presenting the state of the art. Before presenting the FECoM approach, it is necessary to explain what are the limitations of
|
||
existing techniques and what are the challenges that are not addressed yet. Table 2 fairly addresses such points and shows good motivations to support yet another energy measurement technique. However, presenting such a table in Section 6 is too late. Thus, I suggest moving such a table before the approach section and adding descriptive texts introducing and motivating the considered comparison features. I would also include technologies like Monsoon in the table to present a complete view of the field.
|
||
- I would complete RQ3 with a set of criteria that users can consider when she has to select the energy measurement technique that best fits the particular requirement at hand. For instance, if the expected granularity is at the system level, FECoM might not be necessary, or if the acceptable sampling rate is a matter of seconds, CodeCarbon can be a possible choice
|
||
- Concerning RQ1, I like the given summary. However, I'm missing an explicit discussion about using the output. What is the take away message that we can learn and that we can leverage on?
|
||
- The description of Equation (1) needs to be improved. An explanatory example to concretely show the different ingredients of
|
||
the equation would make the paragraph easier to understand.
|
||
- The negative value for RAM energy consumption for some of the considered project is not clear.
|
||
- [[REVIEWS/Notes]]
|
||
- |