Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
(Continuing my answer to "contributions"... I ran out of space) STRONG POINTS/CONTRIBUTIONS 1) The false positive rates and false negative rates observed when using AutoPerf are impressively low. 2) The paper is very well written and easy to follow. 3) The evaluation uses a wide variety of significant, real-world codes. NEGATIVE POINTS 1) The paper lacks a lot of technical depth and novelty… autoencoders for anomaly detection are widely used, and the problem domain (detecting performance bugs) has been studied previously as well. 2) It strikes me that a version of AutoPerf that was production quality could not ignore commit logs, as the authors seem to have done. Knowing what was changed in the code between P_i and P_i+1 could be very, very helpful. 3) The idea of using k-means clustering can speed things a bit, but it is perhaps not the most natural way to do this. Instead, why not just conditions the identity of the function? 4) The evaluation focuses on finding synthetic anomies, rather than real, known performance bugs. DETAILED COMMENTS One comment is that I’m not sure it makes a lot of sense to train separate autoencoders for each function (or group of functions, if you are doing the k-means thing). Likely, there are going to be certain characteristics of the distributions that are shares across all functions, and I worry that you are wasting a lot of compute power by relearning everything. One simple idea would be to learn a single autoencoder, but you would condition the encoder and decoder on the function. That is, you have a separate input to both your encoder and decoder networks that is the identity of the function, but you are conditioning on this (you are not trying to reconstruct the function label), This may allow you to get rid of the multiple autoencoders entirely, and use only one of them. It is interesting that there is this implicit assumption that any changes in performance between P_i and P_i+1 are bad… that they represent depredations in performance. Clearly, this is not the case. Someone can fix a performance bug across versions. Currently, you are likely to flag such a fix as a bug. Is there an easy way to differentiate between good and bad changes? Can this be done without having training data? One big concern that I have about this paper is technical depth and novelty. When it comes right down to it, the authors are simply proposing to use a state-of-the-art approach to anomaly detection to the problem of finding changes in performance across multiple versions of a program. It appears that using anomaly detection to fund performance problems is not new or novel, and I know enough about anomaly detection to know that sing an autoencoder + reconstruction error to do anomaly detection is a fairly standard thing to do nowadays. So what we are left with is a solidly-written paper that applies a standard anomaly detection methodology to a problem that has been studied previously. For this reason, it is tough to get too excited about the paper. It would be nice to know how many functions are in each of the tested codes. I just assumed that a lot of functions were tested for performance bugs, but according to Figure 5, this may not be the case. In fact, I’m surprised that Boost, for example has k = 7 before clustering. Does this mean that Boost has 7 functions?!?! This is very surprising, since Boost is huge. Obviously, there is something that I am missing here, or else maybe the authors have radically down-sampled the codes for evaluation. Regardless, this needs to be explained. It would be good to have more detailed about the anomalies introduced. Note using val anomalies weakens the experimental evaluation, but not having detailed on exactly how the anomies were introduced weakens it even more.
= Originality This is an application paper, proposing a new model/task combination. I am not an expert in the area, but the application of using zero-positive learning with autoencoders to such low-level hardware instrumentation data was an interesting new idea to me. = Clarity The paper is very well-written, and after reading it, I feel that I understood the approach well and could reproduce it. = Quality The article presents a basic idea together with a number of practical extensions (such as the clustering of functions) that make it more applicable. The experiments support most claims, though the number of examples and example runs are low, and choices there are unexplained: Why were 4 four benchmark programs (blackscholes, bodytrack, dedup, streamcluster) run more often for the "normal" version of the program than for the anomalous version? [The authors replied to this saying "Performance anomalies in some applications do not occur for all test cases". I do not understand how that changes the number of executions: Shouldn't the number of test cases actually executed be constant across the two versions of the program?] The phrasing of the metrics could also be misunderstood, as the "false positive" rate is often interpreted as the ratio of wrongly generated warnings in program analysis, whereas here, it related to the overall number of analysed runs. This may be problematic because the system is likely to analyse "normal" programs far more often than those with performance regressions, and so the rate of wrongly reported runs may still be very high. I am also somewhat surprised that considering differing functions alone is sufficient, as many performance regressions stem from a mismatch between allocation and access patterns, but the performance will only be measured (and hence noticed as degreaded) at access time. Hence, in cases where changes to data allocation causes a slowdown at access time, the heuristic of only considering changed functions can easily fail. [The authors replied to this explaining that additional analyses could identify functions that could be impacted by changes; however, I believe that the reply does not consider the implication that such heuristics would mean that much more code would need to be analysed, making it harder to find anomalies] It would be helpful if the authors could comment on these issues in their feedback, and provide some insight if they were observed in the benchmarks. = Significance I believe this submission has little significance for the machine learning community, as it does not put forth a new research direction or insights into ML. If the authors indeed release their implementation and dataset, other researchers may also investigate this task, but I feel that a submission to a conference of the software engineering or performance analysis communities may be more impactful. However, I understand the pain of getting cross-discipline work published, and believe that NeurIPS would be a suitable venue for this work.
--Summary-- This work introduces AutoPerf, a tool for software performance monitoring. This approach attempts to detect regressions (performance bugs) in software by examining changes in hardware telemetry data. In particular, they examine HPCs (Hardware Performace Counters) which contain metrics about low level system events such as cache misses and instruction counts. This can be useful for detecting performance bugs encountered in parallel programming scenarios such as cache contention. The authors train an autoencoder on HPC samples from a non-buggy program and compare the reconstruction error against samples from a candidate program in order to detect whether the candidate program introduces a performance regression. Since the number of functions can be large, the authors suggest clustering similar functions using k-means. The authors evaluate their method across 3 types of regressions, 10 performance bugs and 7 datasets and show that their method outperforms competing methods. --Strengths-- - This is a well written paper that makes an important contribution to an emerging field of ML for systems. - The authors do careful analysis and present a strong case for the value of their method. - This method is likely to be immediately useful as a tool for catching performance regressions in software. --Quality-- Overall this paper is of good research quality and the authors have undertaken careful analysis. The authors can strengthen their paper if they address the following issues: - From a modeling perspective, the authors need to justify why they chose an autoencoder. The main argument put forward by the authors is that supervised data is not available. However, the authors did not consider other unsupervised density modeling techniques such as Gaussian Mixture Models (GMMs) or Variational Auto Encoders (VAEs). These models can be used to directly evaluate the likelihood of test data instead of going through a reconstruction step. - Reconstruction error may not be Gaussian distributed. Looking at figure 3(a) it seems like the distribution is somewhat skewed. The authors may need to study the skew and kurtosis of reconstruction errors in order to strengthen this claim. Alternatively, if they modeled the density directly (for instance using a VAE), then they could apply the “likelihood ratio test” in order to pick deviants. - K-means clustering for scaling up to many functions has tradeoffs. For e.g. Sec 4.3.3 mentions that a function is assigned to a cluster if a majority of its HPC signatures belong to it. This is heavily context dependant. Suppose a test scales O(n) in input size, then differently sized tests will belong to different clusters. - The authors should provide end-to-end timing analysis of their method. Right now it takes 84 minutes for autoencoder training, precluding testing at presubmit time as part of a continuous integration setup. Alternate models such as GMMs can be trained faster. - Sec 5.2 describe how regressions were introduced in the code. However no commentary is made regarding the interestingness of the regression. The authors should specify if they took efforts to break the code in varying and interesting ways. - Minor comment - performance enhancement might be flagged as a bug as well since it causes deviation from expected behavior. Please clarify that deviation is one-sided. --Clarity-- The paper is well written and was easy to follow. A few minor comments: L231 replace x with y for false negative rate. The variable x is overloaded. --Reproducibility-- The authors can aid reproducibility by providing the following (if necessary, in a supplementary section): Details of using PAPI and all the HPCs considered. Details about the hyperparameters used for training their autoencoders. The training procedure and how they split their data for training and validation. --Clarifications/comments-- - The claim in Sec 1, P3 “...AutoPerf, can learn to diagnose potentially any type of regression…” is perhaps too strong. There can be various causes for regressions many of which are not fine-grained enough to be captured by HPCs. - The authors need to address limitations of their method and list causes of false alarms for a candidate program such as (1) inherent randomness (2) unrelated confounding factors. As well as false negatives - AutoPerf is only as strong as the tests. - It’s not clear from the paper, the dimensionality of input to the autoencoder. Does it take in metrics from all 50 HPCs in a vector? If not, then cross-dimensional correlations will be lost.