-
Notifications
You must be signed in to change notification settings - Fork 4
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
JOSS Short Paper Review Comments #11
Comments
Thank you for the detailed review. I will reply inline.
We added the citations for WCET, MLE and PWM and we expanded the acronym. For BM/PoT/GED/GPD we added a couple of books as a general citation for EVT.
Added
Thank you for the detailed explanation. We added a couple of sentences for (b) and the citation, but the in-depth discussion of requirements cannot fit in the very short paper, considering that we should provide also the overall EVT theory basics in case. We emphasized the fact the requirements must be satisfied to obtain a reliable pWCET.
We implemented the CV version of EVT, using the exponential model (that is just a specialized case of the GPD).
This is a very recent work (Apr 2018) and for this reason we decided to do not include directly in
Thank you, reference added.
Fixed. The first names of a couple of references are still missing because they are old articles and we cannot find the original names. Anyway, we provided the DOI for all articles having it. |
Thank you for the paper changes. In general my comments are addressed. Some minor additional changes in order to close this issue:
Regarding the EQMAE I understand your argument, so I agree that you don't have to implement it for the paper to be accepted. How about #12 and #13? If they do not break chronovise's flow, they should be easy to implement and will increase the impact of your software. |
Thank you, all comments fixed.
Ok, PWM is another name for L-moments, so I closed #14 and wrote it in the paper. I implemented GMLE also and closed #13, thank you for suggestions. |
Thank you for the document and code updates, we are almost there :) On the other hands -> On the other hand |
Unfortunately it's not simple to notice them without having the possibility to compile markdown+bibtex locally (or at least I didn't find an easy way to do it). Thank you for detecting them. Fixed 👍 |
Thank you @federeghe for the quick fix. There is a last Sies remaining and a typo (aimis->aims). |
Thanks, fixed |
No paper mentioning WCET is complete without reference to its seminal paper: https://doi.org/10.1145/1347375.1347389 Please add the citation after the first appearence of WCET.
The MLE and PWM abbreviations are mentioned without explaining their meaning. Please provide their definition. Moreover, you need to provide a citations for them, as you have already done for the Kolmogorov-Smirnov and Anderson-Darning tests. A citation is needed also for Block-Maxima, Peak-over-threshold, Generalised Extreme Value and Generalised Pareto Distribution.
Also, seminal papers and state-of-the-art resources for MBPTA are missing:
https://doi.org/10.1109/ECRTS.2012.31 was the first paper to introduce MBPTA, which is based on EVT. Moreover, an appropriate citation for EVT should be provided.
https://www.tdx.cat/handle/10803/460819?locale-attribute=en my PhD thesis explains that in order to safely use MBPTA certain properties needs to be present. This is that:
a) the execution times processed by EVT need to be independent and identically distributed random variables (i.i.d.). This is a requirement coming from EVT theory and has been relaxed later by Santinelli et al under certain circumstances.
b) the events that take place during analysis and affect the execution time of the program (eg. cache misses) must be equal or worse than the ones that take place at operation. This is a requirement of MBPTA in order to ensure that the pWCET is not underestimated, and it is described in the literature as
representativeness.
It is very important to note that those properties are not present in existing platforms, so MBPTA cannot be applied on them without additional measures. In my thesis I proposed hardware and software solutions that make this possible. A simple solution is to make the events that affect the execution time, eg. cache misses, random from execution to execution, so that both properties are satisfied by construction: 1. the cache misses do not depend from one execution to another -> independence, 2. the events follow the same random distribution from execution to execution -> identical distribution, 3. due to identical distribution, the events that happen at analysis are the same with the ones at operation.
Currently the paper only mentions the requirement a) but not b). I think it is very important that the paper mentions both requirements of MBPTA, to prevent that people are using chronovise (and therefore MBPTA) blindly, without ensuring its correct application, because this can result in unsafe pWCET estimates.
https://doi.org/10.1145/3065924 identifies and solves a problem of classical MBPTA, proposing MBPTA-CV; MBPTA-CV takes values only from the tail of the distribution, which are used to model the tail of the distribution with EVT. In contrast, original MBPTA uses values from the entire distribution processed with Block-Maxima or Peak-Over Threashold, which may not belong to the tail of the distribution, resulting in a Generalised Extreme Value (GEV) or Generalised Pareto (GP) distribution with shape that is unsuitable for pWCET use. Thefore, MBPTA-CV can obtain pWCET in cases that original MBPTA fails, while it requires the collection of fewer samples.
https://doi.org/10.1145/3185154 This independent comparison study performed a comparison between the Exponential Model employed by MBPTA-CV and the Block-Maxima or Peak-Over Threashold used by original MBPTA. It shows that the Exponential Model provides sound results, while Block-Maxima or Peak-Over Threashold can provide unsafe pWCETs (underestimation).
In addition to the above missing references, it is worth to note that MBPTA-CV is available as open source from https://doi.org/10.5281/zenodo.1065776 Therefore, the claim in the paper that source code for MBPTA is not available has to be revised. However, I still think that chronovise is a useful tool for MBPTA, because it can provide a common platform for the comparison of the different approaches.
In this aspect, I would like to see MBPTA-CV (Exponential Model and selecting data only from the tail) to be implemented in chronovise, as well as at least the EQMAE (Estimated Quantile's Mean Absolute Error) minimisation criterion proposed in https://doi.org/10.1145/3185154 although the implementation of the
Generalised Maximum Likelihood Estimation (GMLE) and L-moments would be nice to have, too. The reason is that those missing features are considered the current state of the art in MBPTA, and therefore they must be included in chronovise in order to increase its use from the research and engineering community. I believe that implementing those additional features will not be very difficult, since chronovise seems to be extensible.
Finally, please fix some entries in the references section (Ieee, Sies, Rtas) and make sure that all entries follow the same format (some entries have the full first names of the authors while some others only the first letter, some entries have first the surname of the author etc).
The text was updated successfully, but these errors were encountered: