Pith. sign in

REVIEW 4 major objections 4 minor 41 references

Using seven years of production logs from the Theta supercomputer, this paper claims that the forecastability of HPC hardware errors is determined by the temporal structure of the error series—regularly occurring Minor errors are predictabl

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 23:36 UTC pith:OHIDXPTX

load-bearing objection A useful, honest feasibility study on forecasting HPC hardware error logs, but the central Minor-vs-severe contrast may hinge on an unexamined dedup threshold. the 4 major comments →

arxiv 2608.01648 v1 pith:OHIDXPTX submitted 2026-08-03 cs.LG

Evaluating Forecasting Techniques for Hardware Errors on a Large-scale HPC System

Kaiyuan Liao , Xiwei Xuan , Tanwi Mallick , Kevin Brown , Christopher D. Carothers , Kwan-Liu Ma This is my paper
classification cs.LG
keywords high-performance computinghardware error logstime series forecastingdeep learningsystem reliabilityfeasibility analysisTheta supercomputerrare-event prediction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper asks when standard time-series forecasting can predict hardware errors in a large supercomputer, and answers that the deciding factor is the temporal shape of the error series, not the forecasting model. On seven years of logs from the Theta system, correctable Minor errors appear often enough and regularly enough that LSTM and Transformer models, aided by calendar and differencing features, forecast them with moderate accuracy. Intermediate, Major, and Critical errors, by contrast, are long runs of zeros punctuated by rare spikes; every tested model—statistical, recurrent, convolutional, and attention-based—fails to anticipate the spikes, and differences among models mostly reflect noise. The paper therefore positions forecasting not as a universal failure-prediction solution but as a targeted instrument whose usefulness depends on temporal regularity.

Core claim

The paper's central claim is that forecastability is governed by temporal regularity rather than model choice. Using the Theta hardware error dataset, the authors build four daily count series by consolidating raw logs with a 70-second redundancy threshold and ranking error codes into Minor, Intermediate, Major, and Critical severities. One-step-ahead benchmarks show that on the Minor series—which carries 93.6% of all logged errors—LSTM achieves the best errors (MAE 0.05397, MSE 0.00730, RMSE% 30.29), with AR and ARIMA competitive; adding day-of-week, first/second differences, and exponential moving average features further improves LSTM and Transformer but degrades TCN. On Intermediate, Maj

What carries the argument

The argument is carried by four severity-graded daily time series built from the raw Theta logs, and by the contrast between them. After removing exact duplicates and consolidating persistent errors with a 70-second window (following the prior log-analysis procedure referenced as [17]), each error code is mapped to one of four severity levels, producing Minor (93.6% of events), Intermediate, Major, and Critical daily count series. The paper's one-step-ahead forecasting protocol—look-back window of 14 days chosen from the system's biweekly maintenance cycle, sliding-window evaluation, and MSE/MAE/RMSE% metrics—makes the regularity contrast visible: stable series are learnable, sparse series a

Load-bearing premise

The claim that forecasting success depends solely on temporal regularity assumes the 70-second consolidation and four-level severity classification inherited from prior work faithfully reflect the true error process; if another reasonable deduplication threshold or severity mapping makes the severe series look regular, the paper's central contrast would be an artifact of preprocessing.

What would settle it

Re-run the entire experiment on the raw Theta logs with the consolidation threshold varied over a wide range (for instance 10, 70, and 300 seconds) and with alternative severity mappings, then compare the forecasting gap between Minor and severe series: the claim predicts the gap persists for every reasonable preprocessing choice. If some threshold or mapping makes Major or Intermediate errors as forecastable as Minor errors, the paper's conclusion that temporal regularity alone determines forecastability is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • The forecasting gap between Minor and severe series is not a modeling failure: on sparse bursty series, every tested family—AR, ARIMA, LSTM, TCN, Transformer—performs similarly poorly, so gains are unlikely to come from simply swapping models.
  • Domain-informed features (day-of-week, first/second differences, EMA) improve LSTM and Transformer on the predictable Minor series but degrade TCN, so feature engineering must be paired with the right architecture.
  • The best look-back window depends on architecture (LSTM 14, TCN 1, Transformer 10), so the paper's default 14-day window is not universally optimal.
  • Useful forecasting in HPC error analysis is confined to error types with recurring temporal patterns; rare events will need different formulations rather than more one-step-ahead regression.
  • Performance differences on severe series mostly reflect rare events and random variation, so comparisons on such data should be read with caution; the paper's metrics on those series do not indicate genuine pattern learning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper does not test it, but varying the 70-second consolidation threshold (e.g., 10 vs 300 seconds) would directly probe whether the Minor-versus-severe predictability gap is intrinsic to the hardware or an artifact of log preprocessing.
  • Because the severe series are mostly zeros, the paper's negative result implies that treating next-day severe-error occurrence as a binary/probabilistic classification, or predicting time-to-next-burst, could be more informative than count regression; this is left for future work.
  • If the claim generalizes, a practical takeaway the authors leave implicit is that HPC sites should run a quick regularity diagnostic (nonzero-day fraction, autocorrelation at the maintenance period) before choosing a forecasting model.
  • The predictably regular Minor series may itself serve as an early-warning input for severe episodes, since Minor errors often accompany latent instability; the paper records the regularity but does not connect the two signals.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper benchmarks one-step-ahead daily forecasting of hardware error counts from the Theta supercomputer (June 2017–January 2024), using eight models (LAST, Rolling Mean, AR, ARIMA, HA, LSTM, TCN, Transformer). Following the preprocessing of Brown et al. [17], raw logs are deduplicated, consolidated with a 70-second threshold, classified into four severity levels (Minor, Intermediate, Major, Critical), and aggregated into daily time series. The central empirical claim is that forecasting success is governed by temporal structure: Minor errors, which exhibit stable recurring patterns, are predictable — particularly by LSTM and Transformer with engineered temporal features — while Intermediate, Major, and Critical series are sparse and burst-dominated, so that no model achieves meaningful prediction. The paper concludes with feasibility boundaries for forecasting-based HPC error analysis and suggests directions such as probabilistic/anomaly-aware modeling.

Significance. If the empirical contrast holds, this is a useful negative result for HPC prognostics: it demarcates where standard one-step-ahead time-series forecasting adds value and where point-process or anomaly-detection methods are required. The paper is honest about its limitations, uses seven years of production data, cites a publicly available dataset, and includes a look-back-window sensitivity analysis. The main risk is that the central boundary is downstream of preprocessing choices inherited from [17] without validation on Theta, and the quantitative support lacks basic controls relative to trivial baselines and uncertainty quantification. These are fixable within the manuscript's scope.

major comments (4)
  1. [§3.2, §5.2, §6] The 70-second consolidation threshold is a free parameter borrowed from Brown et al. [17] and applied to Theta without validation or sensitivity analysis. The paper's central conclusion — that Intermediate, Major, and Critical series are 'sparse and burst-dominated' and therefore intrinsically unpredictable — is derived from series that have already had bursts collapsed by this threshold. If a shorter threshold (or no consolidation, or counting raw entries) produced daily series with stronger autocorrelation, the feasibility boundary would shift. Please add a sensitivity analysis over the consolidation threshold (e.g., 0, 10, 70, 300, 3600 seconds) and report how the proportion of zeros, lag-1 autocorrelation, and forecasting errors change. Without this, the main claim may describe a preprocessing artifact rather than a property of the hardware error process.
  2. [Table 1, §5.2] For the sparse severe series, no trivial baseline (e.g., constant-zero forecast) or spike-focused metric is reported. Because these series are mostly zeros, MAE and MSE can be small for a model that always predicts zero, while RMSE% is inflated by rare peaks; the observation that 'no method demonstrates clear dominance' is not sufficient evidence that forecasting is inherently limited. Please report a constant-zero predictor and event-level metrics (e.g., spike hit rate, precision/recall on days with nonzero counts, or Prequential score on event days). This would directly test whether any model outperforms the trivial forecast and would substantiate the feasibility-boundary claim.
  3. [Table 1, §5.1, Fig. 5] The paper reports only mean performance over 10 runs, without standard deviations, confidence intervals, or significance tests. The claim that 'LSTM consistently performing best' on the Minor series may be within run-to-run noise, especially since LSTM, Transformer, and ARIMA have very close MAE/MSE values (e.g., 0.05397 vs. 0.05399 vs. 0.05410). Please report variance across runs and apply a paired significance test (Diebold-Mariano or Wilcoxon) between the leading models, at least for the Minor series. Figure 5 shows a scaled ±1 standard deviation band; the unscaled standard deviation should be stated in the text.
  4. [§4, Fig. 4, Table 1] The main comparison in Table 1 uses a unified look-back window Lx=14 for all deep models, but the sensitivity analysis in Fig. 4 shows that TCN performs best at Lx=1 and Transformer at Lx=10, while LSTM is best at Lx=14. The model ranking in Table 1 is therefore not at each model's preferred configuration. Since the abstract and §5.2 emphasize that 'LSTM and Transformer architectures with temporal features' are particularly accurate, the comparison should either use per-model optimal windows or report both the unified-window and tuned-window results. Otherwise the relative advantage of one architecture over another is not cleanly identified.
minor comments (4)
  1. [Table 1] The column header 'R%' is not defined in the caption or text. Based on §5.1 it appears to be RMSE as a percentage of the series mean; please define it explicitly.
  2. [Fig. 5] The caption says the shaded band is a 'scaled ±1 standard deviation' and the text says the band is scaled by a factor of eight. This is confusing; state the actual standard deviation and describe the scaling in the caption.
  3. [§3.1] The sentence 'downtime accounting for approximately 5.33% of operation (179 scheduled and 24 unscheduled events), most lasting less than 24 hours' is imprecise: include the time unit for downtime percentage and clarify whether 'most' refers to scheduled, unscheduled, or all outages.
  4. [Fig. 3] The composition percentages in Fig. 3 are hard to read and their relation to the severity distribution in Fig. 2 is not explained. Add axis labels and a short caption sentence describing how the inner percentages should be interpreted.

Circularity Check

0 steps flagged

No significant circularity: the forecasting conclusions are independent empirical results; the self-citation to [17] supplies preprocessing inputs, not the derived claim.

full rationale

This paper is an empirical benchmark, not a mathematical derivation: there is no equation chain that reduces a claimed prediction to fitted inputs. The only self-citations to [17] provide data-cleaning rules (70-second consolidation) and a severity classification; these are preprocessing inputs, explicitly adopted as prior work, and they are not presented as theorems or uniqueness results. The central claim—Minor errors are forecastable while Intermediate, Major, and Critical are not—is directly evaluated by held-out one-step-ahead forecasting on a chronological test split. No fitted parameter is renamed as a prediction: look-back lengths are either selected from operational periodicity (Lx=14) or varied in a genuine sensitivity analysis (Figure 4), and statistical hyperparameters are grid-searched with best configurations reported, which is standard benchmarking practice rather than a forced prediction. The absence of sensitivity analysis on the 70-second threshold is a validity limitation, but it does not make the argument circular: the conclusions are conditional on the chosen preprocessing, and the paper does not claim to derive the preprocessing from forecasting performance. Self-citation for methodological inheritance is normal and is not load-bearing here, because the cited work does not assert the paper's feasibility-boundary conclusion. Overall, no circular step is identifiable from the text.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

The central empirical claim rests mostly on preprocessing choices from prior work, daily aggregation, and the selection of evaluation metrics. No new theoretical entities or forces are introduced. The main free parameters are hyperparameters chosen by hand or inherited from [17], and the most consequential assumption is that the cleaned daily series faithfully represent hardware error dynamics.

free parameters (8)
  • Look-back window Lx=14 for deep models = 14 days (varied in sensitivity analysis)
    Chosen based on observed biweekly operational periodicity, not fitted to the target; affects all deep model results.
  • EMA span=5 for feature engineering = 5 days
    Hand-selected smoothing span for the Minor series feature set; no sensitivity analysis reported.
  • Rolling mean look-back Lx=3 = 3 days
    Heuristic baseline configuration fixed by the authors.
  • AR look-back Lx=14 = 14 days
    Statistical baseline configuration.
  • ARIMA sliding window Lx=100 = 100 days
    Chosen for numerical stability during retraining at each step.
  • HA seasonal period P=14 = 14 days
    Seasonal heuristic keyed to the biweekly maintenance cycle.
  • 70-second consolidation threshold = 70 seconds
    Inherited from prior work [17]; controls how persistent errors are collapsed and shapes all four series.
  • Chronological train/test split 80/20 = 1934 training / 483 test days approximately
    Evaluation design choice that determines reported metrics.
axioms (5)
  • domain assumption The four-level severity classification of error codes from [17] is a valid grouping for forecasting analysis.
    Invoked in Section 3.2. If category boundaries are wrong, the predictability contrast across severity levels is an artifact.
  • domain assumption Deduplication by exact duplicate removal and 70-second persistent-error consolidation removes artificial correlation without destroying predictive signal.
    Invoked in Section 3.2. A different consolidation threshold could change which series appear regular.
  • domain assumption Daily count aggregation is the appropriate temporal resolution for one-step-ahead HPC error forecasting.
    Used throughout Section 5. Sub-daily or event-level forecasting might behave differently for bursty errors.
  • domain assumption MSE and MAE on normalized daily counts are sufficient to evaluate predictability for HPC error series.
    Used in Section 5.1. On sparse series these metrics are dominated by zero-valued stretches and do not measure spike detection.
  • domain assumption Standard stochastic training with Adam and 10 repeated runs reliably estimates model performance.
    Invoked in Section 5.1. No variance or significance information is reported.

pith-pipeline@v1.3.0-daily-deepseek · 8329 in / 12506 out tokens · 132818 ms · 2026-08-04T23:36:22.262320+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Evaluating Forecasting Techniques for Hardware Errors on a Large-scale HPC System." pith.science (2026). https://pith.science/paper/OHIDXPTX

@misc{pith2026260801648,
  author       = {Pith},
  title        = {Pith review of: Evaluating Forecasting Techniques for Hardware Errors on a Large-scale HPC System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OHIDXPTX}},
  note         = {Machine review of arXiv:2608.01648}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Hardware error logs in high-performance computing (HPC) systems provide early signals of abnormal behavior, yet there remain challenges in effectively forecasting these errors using modern predictive methods. This work investigates the boundaries of applying time series forecasting to HPC hardware error dynamics. We use seven years of production logs from the Theta supercomputer to evaluate the predictive efficacy of classical statistical and deep learning models. Our results show that forecasting effectiveness depends strongly on the temporal structure of the error series: regularly occurring and structurally stable errors can be modeled accurately, particularly by LSTM and Transformer architectures with temporal features, while sparse and burst-dominated errors remain difficult to predict. Rather than proposing a deployment-ready failure prediction framework, this study provides empirical guidance on when forecasting is effective and highlights potential directions for improving forecasting accuracy in HPC hardware error analysis.

Figures

Figures reproduced from arXiv: 2608.01648 by Christopher D. Carothers, Kaiyuan Liao, Kevin Brown, Kwan-Liu Ma, Tanwi Mallick, Xiwei Xuan.

Figure 1
Figure 1. Figure 1: Workflow of the HPC hardware error forecasting: (1) preprocessing raw logs into severity-based daily time series, (2) forecasting using statistical and deep learning models, and (3) evaluation via performance comparison and visualization. runtime overhead at extreme scale [5]. Consequently, there is growing interest in failure avoidance strategies [6], which aim to identify early indicators of abnormal beh… view at source ↗
Figure 2
Figure 2. Figure 2: Severity distribution of hardware errors. duplicated and persistent records where identical failures are repeatedly logged within short intervals, introducing artificial correlation that must be removed before constructing forecasting-ready time series. 3.2 Data Preprocessing Redundancy Removal. We follow the cleaning procedures described in [17]. Two forms of redundancy are present. First, exact duplicate… view at source ↗
Figure 3
Figure 3. Figure 3: Daily aggregated hardware error counts in Theta (2017–2024) across four sever￾ity levels with error-type composition. Minor errors show sustained activity, while In￾termediate, Major, and Critical errors become sparse and spike-dominated. For deep learning models, we adopt a unified look-back window of Lx = 14, selected based on observed operational periodicity in the system. Statistical models are evaluat… view at source ↗
Figure 4
Figure 4. Figure 4: Impact of look-back window size Lx on forecasting performance (MSE) [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: LSTM predictions on the Minor Error series. The orange line denotes ground￾truth counts, and the blue line shows the mean prediction across 10 runs. The shaded band represents a scaled ±1 standard deviation, indicating prediction variability. Feature Engineering on Minor Error Data. To further improve forecasting performance on the Minor Error series, we augment the input with additional temporal features.… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages

  1. [1]

    Failover strategy for fault tolerance in cloud computing environment,

    B. Mohammed, M. Kiran, K. M. Maiyama, M. M. Kamala, and I.-U. Awan, “Failover strategy for fault tolerance in cloud computing environment,” Software: Practice and Experience, vol. 47, no. 9, pp. 1243–1274, 2017

    work page 2017

  2. [2]

    Improving the computing efficiency of HPC systems using a combi- nation of proactive and preventive checkpointing,

    M. S. Bouguerra, A. Gainaru, L. Bautista Gomez, F. Cappello, S. Matsuoka, and N. Maruyama, “Improving the computing efficiency of HPC systems using a combi- nation of proactive and preventive checkpointing,” in 2013 IEEE 27th International Symposium on Parallel and Distributed Processing, 2013, pp. 501–512

    work page 2013

  3. [3]

    Desh:deeplearningforsystemhealth prediction of lead times to failure in HPC,

    A.Das,F.Mueller,C.Siegel,andA.Vishnu,“Desh:deeplearningforsystemhealth prediction of lead times to failure in HPC,” in Proceedings of the 27th International Symposium on High-Performance Parallel and Distributed Computing, 2018, pp. 40–51

    work page 2018

  4. [4]

    Automatic recognition of intermittent failures: an experimental study of field data,

    R. K. Iyer, L. T. Young, and P. V. K. Iyer, “Automatic recognition of intermittent failures: an experimental study of field data,” IEEE Transactions on Computers, vol. 39, no. 4, pp. 525–537, 1990

    work page 1990

  5. [5]

    Fault prediction under the microscope: a closer look into HPC systems,

    A. Gainaru, F. Cappello, M. Snir, and W. Kramer, “Fault prediction under the microscope: a closer look into HPC systems,” in SC’12: Proceedings of the Inter- national Conference on High Performance Computing, Networking, Storage and Analysis, 2012, pp. 1–11

    work page 2012

  6. [6]

    Checkpointing vs. migration for post- petascale supercomputers,

    F. Cappello, H. Casanova, and Y. Robert, “Checkpointing vs. migration for post- petascale supercomputers,” in 2010 39th International Conference on Parallel Pro- cessing, 2010, pp. 168–177

    work page 2010

  7. [7]

    A meta-learning failure pre- dictor for Blue Gene/L systems,

    P. Gujrati, Y. Li, Z. Lan, R. Thakur, and J. White, “A meta-learning failure pre- dictor for Blue Gene/L systems,” in 2007 International Conference on Parallel Processing (ICPP 2007), 2007, pp. 40–40

    work page 2007

  8. [8]

    G. E. P. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung,Time Series Analysis: Forecasting and Control. John Wiley & Sons, 2015

    work page 2015

  9. [9]

    R. G. Brown,Smoothing, Forecasting and Prediction of Discrete Time Series. Courier Corporation, 2004

    work page 2004

  10. [10]

    A comparison of ARIMA and LSTM in forecasting time series,

    S. Siami-Namini, N. Tavakoli, and A. S. Namin, “A comparison of ARIMA and LSTM in forecasting time series,” in 2018 17th IEEE International Conference on Machine Learning and Applications (ICMLA), 2018, pp. 1394–1401

    work page 2018

  11. [11]

    Financial time series forecast- ing with deep learning: a systematic literature review: 2005–2019,

    O. B. Sezer, M. U. Gudelek, and A. M. Ozbayoglu, “Financial time series forecast- ing with deep learning: a systematic literature review: 2005–2019,” Applied Soft Computing, vol. 90, p. 106181, 2020. Forecasting Hardware Errors in a HPC System 11

    work page 2005

  12. [12]

    Temporal convolutional neural (TCN) network for an effective weather forecasting using time-series data from the local weather station,

    P. Hewage, A. Behera, M. Trovati, E. Pereira, M. Ghahremani, F. Palmieri, and Y. Liu, “Temporal convolutional neural (TCN) network for an effective weather forecasting using time-series data from the local weather station,” Soft Computing, vol. 24, no. 21, pp. 16453–16482, 2020

    work page 2020

  13. [13]

    Traffic flow forecast through time series analysis based on deep learning,

    J. Zheng and M. Huang, “Traffic flow forecast through time series analysis based on deep learning,” IEEE Access, vol. 8, pp. 82562–82570, 2020

    work page 2020

  14. [14]

    Application of multivariate time-series model for high performance computing (HPC) fault prediction,

    X. Pei, M. Yuan, G. Mao, and Z. Pang, “Application of multivariate time-series model for high performance computing (HPC) fault prediction,” PLOS ONE, vol. 18, no. 10, p. e0281519, 2023

    work page 2023

  15. [15]

    Surrogate modeling for HPC application iteration times forecasting with network features,

    X. Xu, K. A. Brown, T. Mallick, X. Wang, E. Cruz-Camacho, R. B. Ross, C. D. Carothers, Z. Lan, and K. Shu, “Surrogate modeling for HPC application iteration times forecasting with network features,” in Proceedings of the 38th ACM SIGSIM Conference on Principles of Advanced Discrete Simulation, 2024, pp. 93–97

    work page 2024

  16. [16]

    A hybrid hierarchical time series model for predicting HPC job runtime,

    Y. Tian, X. Zhang, F. Yang, W. Yang, G. Xian, J. Yu, and L. Lai, “A hybrid hierarchical time series model for predicting HPC job runtime,” in 2025 10th In- ternational Conference on Computer and Communication System (ICCCS), 2025, pp. 137–142

    work page 2025

  17. [17]

    Analyzing a lifetime of failures on a Cray XC40 supercomputer,

    K. A. Brown, T. Mallick, Z. Lan, R. B. Ross, and C. D. Carothers, “Analyzing a lifetime of failures on a Cray XC40 supercomputer,” in Proceedings of the Cray User Group, 2025, pp. 103–114

    work page 2025

  18. [18]

    Understanding HPC application I/O behavior using system level statistics,

    A. K. Paul, O. Faaland, A. Moody, E. Gonsiorowski, K. Mohror, and A. R. Butt, “Understanding HPC application I/O behavior using system level statistics,” in 2020 IEEE 27th International Conference on High Performance Computing, Data, and Analytics (HiPC), 2020, pp. 202–211

    work page 2020

  19. [19]

    Characterization and identification of HPC applications at leadership computing facility,

    Z. Liu, R. Lewis, R. Kettimuthu, K. Harms, P. Carns, N. Rao, I. Foster, and M. E. Papka, “Characterization and identification of HPC applications at leadership computing facility,” in Proceedings of the 34th ACM International Conference on Supercomputing, 2020, pp. 1–12

    work page 2020

  20. [20]

    Design and implementation of I/O performance prediction scheme on HPC systems through large-scale log analysis,

    S. Kim, A. Sim, K. Wu, S. Byna, and Y. Son, “Design and implementation of I/O performance prediction scheme on HPC systems through large-scale log analysis,” Journal of Big Data, vol. 10, no. 1, p. 65, 2023

    work page 2023

  21. [21]

    A machine learn- ing approach to online fault classification in HPC systems,

    A. Netti, Z. Kiziltan, O. Babaoglu, A. Sîrbu, and A. Bartolini, “A machine learn- ing approach to online fault classification in HPC systems,” Future Generation Computer Systems, vol. 110, pp. 1009–1022, 2020

    work page 2020

  22. [22]

    Characterizing HPC performance variation with monitoring and unsupervised learning,

    G. Ozer, A. Netti, D. Tafani, and M. Schulz, “Characterizing HPC performance variation with monitoring and unsupervised learning,” in International Conference on High Performance Computing, 2020, pp. 280–292

    work page 2020

  23. [23]

    Aarohi: making real-time node failure pre- diction feasible,

    A. Das, F. Mueller, and B. Rountree, “Aarohi: making real-time node failure pre- diction feasible,” in 2020 IEEE International Parallel and Distributed Processing Symposium (IPDPS), 2020, pp. 1092–1101

    work page 2020

  24. [24]

    Runtime level failure detection andpropagationinHPCsystems,

    D. Zhong, A. Bouteiller, X. Luo, and G. Bosilca, “Runtime level failure detection andpropagationinHPCsystems,” inProceedingsofthe26thEuropeanMPIUsers’ Group Meeting, 2019, pp. 1–11

    work page 2019

  25. [25]

    Failure prediction using ma- chine learning in a virtualised HPC system and application,

    B. Mohammed, I. Awan, H. Ugail, and M. Younas, “Failure prediction using ma- chine learning in a virtualised HPC system and application,” Cluster Computing, vol. 22, no. 2, pp. 471–485, 2019

    work page 2019

  26. [26]

    Time machine: generative real-time model for failure (and lead time) prediction in HPC systems,

    K. A. Alharthi, A. Jhumka, S. Di, L. Gui, F. Cappello, and S. McIntosh-Smith, “Time machine: generative real-time model for failure (and lead time) prediction in HPC systems,” in 53rd Annual IEEE/IFIP International Conference on Depend- able Systems and Networks (DSN), 2023, pp. 508–521. 12 K. Liao et al

    work page 2023

  27. [27]

    ARIMA model for accurate time series stocks forecasting,

    S. Khan, “ARIMA model for accurate time series stocks forecasting,” International Journal of Advanced Computer Science and Applications, 2020

    work page 2020

  28. [28]

    Forecasting with a combined model of ETS and ARIMA,

    J. Oh and B. Seong, “Forecasting with a combined model of ETS and ARIMA,” Communications for Statistical Applications and Methods, vol. 31, no. 1, pp. 143– 154, 2024

    work page 2024

  29. [29]

    Deep learning for time series forecasting: A survey of recent advances,

    K. Liao, X. Xuan, and K.-L. Ma, “Deep learning for time series forecasting: A survey of recent advances,”Frontiers of Computer Science, vol. 20, no. 11, Art. no. 2011359, 2026

    work page 2026

  30. [30]

    TimesNet: temporal 2D- variation modeling for general time series analysis,

    H. Wu, T. Hu, Y. Liu, H. Zhou, J. Wang, and M. Long, “TimesNet: temporal 2D- variation modeling for general time series analysis,” in The Eleventh International Conference on Learning Representations, 2023

    work page 2023

  31. [31]

    TVNet: a novel time series analysis method based on dy- namic convolution and 3D-variation,

    C. Li, M. Li, and R. Diao, “TVNet: a novel time series analysis method based on dy- namic convolution and 3D-variation,” in The Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  32. [32]

    Locally connected echo state networks for time series forecasting,

    F. Matzner and F. Mráz, “Locally connected echo state networks for time series forecasting,” in The Thirteenth International Conference on Learning Representa- tions, 2025

    work page 2025

  33. [33]

    Unlocking the power of LSTM for long term time series forecasting,

    Y. Kong, Z. Wang, Y. Nie, T. Zhou, S. Zohren, Y. Liang, P. Sun, and Q. Wen, “Unlocking the power of LSTM for long term time series forecasting,” in Proceed- ings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 11, 2025, pp. 11968–11976

    work page 2025

  34. [34]

    Cross- GNN: confronting noisy multivariate time series via cross interaction refinement,

    Q. Huang, L. Shen, R. Zhang, S. Ding, B. Wang, Z. Zhou, and Y. Wang, “Cross- GNN: confronting noisy multivariate time series via cross interaction refinement,” in NeurIPS, 2023

    work page 2023

  35. [35]

    Irregular multivariate time series forecasting: a transformable patching graph neural networks approach,

    W. Zhang, C. Yin, H. Liu, X. Zhou, and H. Xiong, “Irregular multivariate time series forecasting: a transformable patching graph neural networks approach,” in Forty-first International Conference on Machine Learning, 2024

    work page 2024

  36. [36]

    A time series is worth 64 words: long-term forecasting with transformers,

    Y. Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: long-term forecasting with transformers,” in The Eleventh International Conference on Learning Representations, 2023

    work page 2023

  37. [37]

    CARD: channel aligned robust blend transformer for time series forecasting,

    X. Wang, T. Zhou, Q. Wen, J. Gao, B. Ding, and R. Jin, “CARD: channel aligned robust blend transformer for time series forecasting,” in The Twelfth International Conference on Learning Representations, 2024

    work page 2024

  38. [38]

    Timer-XL: long-context trans- formers for unified time series forecasting,

    Y. Liu, G. Qin, X. Huang, J. Wang, and M. Long, “Timer-XL: long-context trans- formers for unified time series forecasting,” in The Thirteenth International Con- ference on Learning Representations, 2025

    work page 2025

  39. [39]

    TEMPO: prompt-based generative pre-trained transformer for time series forecasting,

    D. Cao, F. Jia, S. O. Arik, T. Pfister, Y. Zheng, W. Ye, and Y. Liu, “TEMPO: prompt-based generative pre-trained transformer for time series forecasting,” in The Twelfth International Conference on Learning Representations, 2024

    work page 2024

  40. [40]

    Time-LLM: time series forecasting by reprogram- ming large language models,

    M. Jin, S. Wang, L. Ma, Z. Chu, J. Y. Zhang, X. Shi, P.-Y. Chen, Y. Liang, Y.-F. Li, S. Pan, and Q. Wen, “Time-LLM: time series forecasting by reprogram- ming large language models,” in The Twelfth International Conference on Learning Representations, 2024

    work page 2024

  41. [41]

    Smart predictive maintenance for high-performance computing systems: a literature review,

    A. L. da Cunha Dantas Lima, V. M. Aranha, C. J. de Lima Carvalho, and E. G. Sperandio Nascimento, “Smart predictive maintenance for high-performance computing systems: a literature review,” The Journal of Supercomputing, vol. 77, no. 11, pp. 13494–13513, 2021

    work page 2021