{"id":"76f501b3-3403-4269-aebc-9104ea0d7881","arxiv_id":"2501.17796","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An incremental variant of multiresolution dynamic mode decomposition (I-mrDMD) updates supercomputer log analyses in seconds rather than recomputing from scratch, with two case studies on the Theta supercomputer.","lead":"This paper presents a faster way to process supercomputer sensor data by updating pattern calculations incrementally as new readings arrive, rather than recomputing everything from scratch. It demonstrates the approach on logs from the Theta supercomputer, where updates complete in seconds, and pairs it with an interactive rack-level visualization.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"I-mrDMD updates only level-1 SVD; higher-level modes stay stale, and the paper never reports a head-to-head accuracy comparison against batch mrDMD on the combined data, so the streaming-fidelity claim is unverified.","rationale":"The reader's weakest-assumption analysis correctly identifies that I-mrDMD is not yet a full incremental mrDMD: only the level-1 SVD is updated, and levels 2-L are deferred. My stress-test confirms this is the load-bearing issue. An online algorithm that ignores the newly arriving high-frequency content cannot claim to produce an accurate multiresolution decomposition of the streaming data unless strong evidence shows the omitted levels are negligible for the target analyses. The paper acknowledges error accumulation over months to years, yet the central use case is monitoring over long time scales, so this is not a peripheral limitation. The reported speedups are credible and the engineering contribution is real, but the accuracy comparison that would justify the headline claim is missing: the paper never directly compares I-mrDMD against batch mrDMD on the same combined data, and it reports no repeated-update experiment. The Frobenius norms in the case studies are absolute values that are hard to interpret without knowing the data scale or the batch mrDMD error on the same data. Therefore the current evidence supports a conditional acceptance: the method is fast, but its fidelity as an incremental mrDMD is unverified. My recommendation is UNCHANGED because my analysis reinforces the reader's CONDITIONAL verdict rather than moving it to a different category.","tokens_in":18095,"tokens_out":5614,"duration_ms":58106,"concrete_test":"Run a controlled comparison on the two datasets from Section V. For the same initial T and appended T1, compute (a) batch mrDMD on the full T+T1 matrix with identical parameters, and (b) I-mrDMD as implemented (level-1 incremental SVD update, levels 2-L left stale). Compare the reconstructed matrices blockwise: compute normalized Frobenius errors ||X_full_recon - X_incr_recon||_F / ||X_full_recon||_F for the old block, the appended block, and the full timeline. Also report batch mrDMD's own reconstruction error on the full data to interpret the 3958/3423 numbers. Finally, repeat the update 10-20 times with 1,000-point increments and track the normalized error as a function of update count.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that I-mrDMD generates mrDMD modes incrementally for online data. The implementation, however, only incrementally updates the level-1 SVD; updates to levels 2-L are explicitly deferred (Section III-A1: \"we leave this step for future work\"; Section VI: \"left as a part of future work\"). For new time steps, the residual after subtracting the new level-1 slow modes is never decomposed into higher-level frequency bands, so the high-frequency part of the appended data is absent from the updated representation unless the old higher-level modes happen to apply. The paper asserts that reconstruction error rises by only \"a sum of 10-5000\" per update and that this is small, but it reports no direct comparison between I-mrDMD and ordinary mrDMD on the same T+T1 data, and no experiment with repeated updates to support the claim that error does not accumulate to a significant level. The reconstruction norms in Section V (3958.58, 3423.847) are absolute and are not compared with batch mrDMD's reconstruction error on the same data, nor with the data norm; they therefore do not establish fidelity. Because Q2 explicitly asks for the accuracy difference between online and regular mrDMD, and the paper never answers it quantitatively, the streaming-accuracy claim is the load-bearing unsupported step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents I-mrDMD, an incremental version of multiresolution dynamic mode decomposition, with the stated goal of enabling fast, accurate analysis of streaming time-series data from HPC system logs. The authors integrate an incremental SVD update at the first mrDMD level, report wall-clock timings for appending new time points on Theta environment logs and Polaris GPU temperature data, and combine the resulting modes with z-score baselines and a D3-based rack visualization to align environment-, job-, and hardware-log behavior in two case studies. The manuscript explicitly states that updates to mrDMD levels 2 through L are left to future work, and it reports reconstruction errors only as absolute Frobenius norms without a batch mrDMD baseline.","tokens_in":18396,"tokens_out":4125,"duration_ms":41576,"significance":"If the central claim holds, an incremental mrDMD that can refresh a multiresolution decomposition quickly enough for streaming sensor data would be practically valuable for HPC monitoring, where environment logs accumulate terabytes per day. The paper's strengths include the use of real-world supercomputer datasets, concrete wall-clock measurements averaged over 10 runs, a public code repository, and an explicit discussion of the algorithm's current limitations. However, the load-bearing assertions about fidelity and about correlation with hardware/job events are not quantitatively established: the implementation updates only the level-1 SVD, and the reported accuracy measures are not compared with batch mrDMD on the same data. The paper is therefore a promising proof-of-concept whose central streaming-accuracy claim needs additional verification before publication.","major_comments":[{"comment":"The central claim of an incremental mrDMD is only partially implemented: the incremental SVD update is applied to level 1 only, and the text explicitly defers updates to levels 2 through L to future work (Section III-A1: \"we leave this step for future work\"; Section VI: \"left as a part of future work\"). For each appended block T1, the residual after subtracting the newly updated level-1 slow modes is not decomposed into higher-level frequency bands, so the high-frequency content of the appended segment is absent from the updated representation unless the old higher-level modes happen to capture it. Because Q1 and Q2 are about the fidelity of the online decomposition, the manuscript needs a head-to-head accuracy comparison between I-mrDMD and batch mrDMD on the same T+T1 data, including reconstruction norms and mode-level differences, and at least one multi-step update experiment to show how error accumulates over repeated appends.","section":"III-A1 and Algorithm 1"},{"comment":"The quantitative support for accuracy is incomplete. The Frobenius norms reported in Section V (3958.58 and 3423.847) are absolute residuals between the actual and I-mrDMD-reconstructed data, but they are not compared with the corresponding batch mrDMD residual on the same data, nor with the norm of the data itself, so they do not by themselves establish that the incremental result is accurate. Likewise, the assertion in Section III-A1 that the reconstruction difference increases \"only by a sum of 10-5000\" per update is not accompanied by an experiment, a table, or a description of how this quantity was measured, for which dataset, or over how many updates. Please provide this comparison to answer Q2 quantitatively.","section":"Section V and III-A1"},{"comment":"The claimed correlation between environment-log patterns and hardware/job failures is supported only by visual inspection of rack views (Figs. 4 and 6). No quantitative measure is reported linking z-scores or mode amplitudes to memory errors or job failures; the baselines and z-score thresholds are manually selected (e.g., 46-57 °C in Case Study 1 and different ranges in Case Study 2). To support Q3, report a quantitative association (e.g., precision/recall, a confusion matrix, or a correlation coefficient) between anomalous z-scores and recorded hardware or job events, and address sensitivity to the chosen baselines and thresholds.","section":"Section V, Case Studies 1 and 2"},{"comment":"The performance comparisons (14.728 s vs 80.580 s; 29.945 s vs 59.263 s; Table I) are for the partial update that skips levels 2-L. Since the paper states that updating levels 2 through L is future work, the reported speedups are for only a subset of the full mrDMD update. If the full update is intended to run online, the paper should either implement and time it, or state clearly that the speedup applies only to the level-1 refresh and that the remaining levels are updated asynchronously with unknown latency.","section":"Section IV"}],"minor_comments":[{"comment":"The notation is inconsistent: the conjugate transpose is written as V' in Eq. (2) but the text says the symbol is the conjugate transpose, and Eq. (6) uses aaa and ai(0) without defining how the initial amplitudes are computed. Please make the notation consistent and add a sentence defining the amplitudes.","section":"Section III-A, Eqs. (2)-(6)"},{"comment":"The alignment specification text says the row and column alignment takes numbers \"-1, 1, 2\" for right-to-left, left-to-right, and bottom-to-top, but then lists \"2 for bottom-to-top\" after already including 2; this appears to be a typo and the mapping from numbers to alignments should be clarified.","section":"Section III-B"},{"comment":"Reference [51] is labeled as the Theta supercomputer but the URL points to the Polaris page at Argonne; either the reference or the URL is incorrect and should be fixed.","section":"References"},{"comment":"Algorithm 1, line 8, uses \"node level\" where the intended variable is likely \"level,\" and the loop \"for previous nodes = 1,2,...L\" is unclear; please rewrite the pseudocode to match the level notation used in the text.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on the authors' earlier work [2], [3] for the baseline mrDMD methodology and for the choice of sampling rate, and the current paper's algorithmic novelty is the incremental SVD integration at level 1 plus the visualization and case studies. This is a legitimate contribution, but the editor should ask the authors to clearly delimit what is new relative to [2]. The biggest risk is that the streaming-fidelity claim remains unverified; the missing batch-vs-incremental comparison should be a required revision rather than an optional extension."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the engineering is real, the headline is too big, and the missing comparison is exactly the one the authors said they would make.\n\nWhat's new: they apply incremental SVD to the level-1 SVD of mrDMD and measure it on two HPC datasets. Updating 5,000 time points takes about 15-30 s versus about 59-80 s for full recomputation, and the scaling table shows partial fits staying roughly flat while batch fits grow. The D3 rack visualization in Jupyter is a practical contribution for operations, and the code is public. Those results are concrete and credible: wall-clock times are averaged over 10 runs, and the update fits inside a 10-30 s sensor window in the tested cases.\n\nSoft spots, in order of size.\n\nFirst, 'incremental mrDMD' overstates what is implemented. Only the level-1 SVD is updated; updates to levels 2-L are deferred (Sections III-A1 and VI). The authors say this plainly, but the abstract and contribution list do not. With stale higher-level modes, the representation is not a fresh mrDMD of the combined stream.\n\nSecond, Q2 asks for the accuracy difference between online and regular mrDMD, and the paper never answers it with numbers. The reconstruction norms (3958 and 3424) are absolute; there is no batch mrDMD run on the same T+T1 data, no data norm for context, and no repeated-update experiment, even though the authors note error can accumulate over months. 'A sum of 10-5000' is not a measurable claim as stated.\n\nThird, windowed mrDMD [38] and streaming DMD [40-42] are cited but not compared. A runtime and accuracy comparison against at least one of these would sharpen the novelty claim.\n\nFourth, the fault-correlation case studies are anecdotal. Fine for a visual analytics demo, but the language sometimes implies more than the evidence shows.\n\nI agree with the reader's conditional verdict. The speedups are real; the streaming-fidelity claim is not yet established. The fix is straightforward: a head-to-head I-mrDMD versus mrDMD on the same appended data, with repeated updates and relative reconstruction errors. That is a revision, not a rewrite.\n\nFor whom: HPC operations teams and visual analytics researchers will get direct value from the rack view and the streaming pipeline. DMD methodologists should wait for the accuracy comparison.\n\nRecommendation: send it to peer review. It deserves a serious referee; the authors shipped code, data, and honest limitation statements. I would cite it for the engineering result while noting the accuracy gap.","headline":"Real speedups from incrementally updating level-1 SVD of mrDMD, but the streaming-accuracy claim is unverified because levels 2-L stay stale and the head-to-head comparison against batch mrDMD is missing.","tokens_in":18946,"tokens_out":4748,"would_cite":true,"duration_ms":44099,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Incremental multiresolution DMD (I-mrDMD) updates supercomputer log analysis in 15–30 seconds instead of 60–80 seconds, making online monitoring practical.","keywords":["high performance computing","incremental analysis","online analysis","multiresolution dynamic mode decomposition","mrDMD","visualization","streaming data","singular value decomposition"],"falsifier":"Run I-mrDMD on a year of streaming sensor data, and after each update recompute the full mrDMD from scratch and compare the two reconstructions. If the Frobenius-norm difference grows with each update rather than staying bounded—or the z-score baselines drift enough to change which nodes are flagged—the claim that the partial update stays faithful over long streams is refuted.","tokens_in":17877,"feed_emoji":"⚡","tokens_out":12101,"duration_ms":98656,"temperature":0.7,"pith_summary":"Supercomputers generate terabytes per day of sensor readings—temperatures, voltages, fan speeds—too voluminous to analyze fast enough to catch problems in real time. This paper claims that an incremental version of multiresolution dynamic mode decomposition (I-mrDMD) can update an existing mode decomposition when new time points arrive, rather than recomputing it from scratch. In tests on real log data from the Theta and Polaris supercomputers, an update takes roughly 15–30 seconds instead of 60–80 seconds, and the reconstructed data stays close to the raw readings (Frobenius norm around 3,400–3,960). The paper pairs the decomposition with z-scores against baseline behavior and a rack-level visualization to expose node states such as overheating, idle nodes, and hardware errors, arguing that this makes online analysis of HPC monitoring data practical.","feed_headline":"Cut supercomputer-log analysis from minutes to seconds","feed_subtitle":"Incrementally updated multiresolution DMD keeps streaming analysis fast enough to catch failures early.","key_machinery":"I-mrDMD is the machinery that carries the argument. It combines multiresolution dynamic mode decomposition (mrDMD)—which recursively subtracts slower-frequency modes from time-series data at each of several scales—with an incremental singular value decomposition update at level 1, so newly arriving time steps extend the existing SVD instead of triggering a complete recomputation. The updated level-1 modes are used to reconstruct and subtract the slow dynamics; the residual is then split and processed at finer levels exactly as in the ordinary mrDMD. Mode power and frequency come from the DMD eigenvalues, and z-scores computed against a baseline set of modes flag nodes whose behavior deviates from normal. The authors note that updates to levels 2 through L are left for future work, and they treat the level-1 update as the main computational saving.","core_discovery":"The paper's central claim is that the multiresolution dynamic mode decomposition (mrDMD), which converts high-dimensional time series into spatiotemporal modes at multiple frequency scales, can be made incremental with little loss of fidelity, enabling online analysis of high-velocity sensor streams. The proposed I-mrDMD algorithm updates the singular value decomposition at the first (slowest) level of the mrDMD hierarchy using an incremental SVD step, recomputes the DMD modes for that level, and then carries the reconstructed slow dynamics through the remaining levels as in ordinary mrDMD. The authors report that on Theta environment logs (4,392 nodes, 50,000 time points), adding 5,000 new time points takes 14.7 seconds rather than 80.6 seconds for a full recalculation; on Polaris GPU temperature data, the update takes 29.9 seconds versus 59.3 seconds. Reconstruction quality, measured by the Frobenius norm of the difference between actual and reconstructed data, is roughly 3,400–3,960 in the two case studies, which the authors present as evidence that the incremental modes capture the underlying system dynamics. They further claim that z-scores of mode-power deviations from baselines separate normal from anomalous node behavior and align with events in the hardware and job logs.","pith_inferences":["Extending the incremental SVD update to the deeper mrDMD levels, which the paper calls embarrassingly parallel, would remove the remaining error accumulation and is a natural next step beyond the paper's level-1-only update.","The 10–5,000 per-update error growth implies an operator could schedule a full recomputation before error crosses a chosen threshold; the paper does not derive such a schedule, but its measurements are enough to build one.","The cleaner separation of baseline from non-baseline readings that the authors observe for mrDMD and I-mrDMD, compared with PCA, UMAP, and t-SNE, suggests the multiresolution frequency decomposition itself—not just the speedup—is what makes the anomaly signal clear; this could be probed on other high-velocity multivariate streams."],"forward_implications":["Update time drops below the sensor sampling interval, so I-mrDMD can keep pace with streaming data from current supercomputers that sample at 0.03–10 Hz.","The decomposition compresses terabytes of environment logs into a small set of spatiotemporal modes, shrinking the data to a size that can be inspected and visualized interactively.","Per-node z-scores against baselines give operators a direct visual signal for overheating nodes, idle nodes, and nodes with persistent hardware errors, aligned across environment, hardware, and job logs.","Because the mode update is data-driven and the rack visualization is parameterized by a layout string, the same pipeline transfers to other large-scale systems without retraining.","The reported reconstruction error grows by roughly 10–5,000 per update, so for streams of weeks or months the incremental result is accurate; for multi-year streams the paper notes error may accumulate and require asynchronous full recomputation."],"supporting_citations":[{"why":"Supplies the incremental SVD update used to extend the level-1 decomposition with new time steps.","marker":"[46]"},{"why":"Defines the mrDMD algorithm that I-mrDMD makes incremental.","marker":"[33]"},{"why":"The previous mrDMD-on-HPC pipeline and baseline z-score approach this work extends.","marker":"[2]"},{"why":"Provides the DMD formulation and reconstruction equation used to compute modes and project the time series.","marker":"[47]"},{"why":"The singular-value hard threshold used to choose the reduced rank of the SVD.","marker":"[49]"},{"why":"Supplies the z-score baseline procedure for flagging deviations in mode power from normal behavior.","marker":"[1]"},{"why":"The windowed mrDMD streaming approach the paper contrasts with, arguing its incremental SVD avoids overlapping-window staggering.","marker":"[38]"}],"fun_headline_variants":["Incremental DMD speeds up supercomputer log analysis from 80s to 15s","Streaming mode decomposition cuts log analysis from minutes to seconds","Incremental multiresolution DMD enables real-time HPC monitoring","I-mrDMD online decomposition for petabyte-scale system logs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that updating only the first (slowest) level of the multiresolution decomposition, while leaving the finer levels' modes unchanged, still gives a faithful picture of the system, even though the paper's own measurements show small reconstruction errors (10–5,000 per update) that can accumulate over months or years of streaming.","fun_headline_variants_meta":{"raw":{"variants":["Incremental DMD speeds up supercomputer log analysis from 80s to 15s","Streaming mode decomposition cuts log analysis from minutes to seconds","Incremental multiresolution DMD enables real-time HPC monitoring","I-mrDMD online decomposition for petabyte-scale system logs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000799,"raw_usage":{"total_tokens":3519,"prompt_tokens":954,"completion_tokens":2565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":2485}},"tokens_in":570,"tokens_out":2565,"duration_ms":18067,"temperature":1.0,"reasoning_tokens":2485,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:31:51.051283+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run I-mrDMD on a year of streaming sensor data, and after each update recompute the full mrDMD from scratch and compare the two reconstructions. If the Frobenius-norm difference grows with each update rather than staying bounded—or the z-score baselines drift enough to change which nodes are flagged—the claim that the partial update stays faithful over long streams is refuted.","supporting_citations":[{"cited_title":"An incremental singular value decomposition approach for large-scale spatially parallel & distributed but temporally serial data – applied to technical flows,","cited_arxiv_id":null,"evidence_quote":"Supplies the incremental SVD update used to extend the level-1 decomposition with new time steps."},{"cited_title":"A Multi-Level, Multi-Scale Visual Analytics Approach to Assessment of Multifidelity HPC Systems","cited_arxiv_id":"2306.09457","evidence_quote":"The previous mrDMD-on-HPC pipeline and baseline z-score approach this work extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DMD formulation and reconstruction equation used to compute modes and project the time series."},{"cited_title":"The optimal hard threshold for singular values is 4 / √ 3,","cited_arxiv_id":null,"evidence_quote":"The singular-value hard threshold used to choose the reduced rank of the SVD."},{"cited_title":"Extracting spatial–temporal coherent patterns in large-scale neural recordings using dynamic mode decomposition,","cited_arxiv_id":null,"evidence_quote":"Supplies the z-score baseline procedure for flagging deviations in mode power from normal behavior."},{"cited_title":"Novel windowed multi- resolution dynamic mode decomposition (wmrdmd): Application to unsteady surface pressure over a wing in flutter,","cited_arxiv_id":null,"evidence_quote":"The windowed mrDMD streaming approach the paper contrasts with, arguing its incremental SVD avoids overlapping-window staggering."}],"review_version":1}