Pith. sign in

REVIEW 4 major objections 4 minor 44 references

TELLER: Non-intrusive Cross-Layer Root-Cause Analysis for LLM Inference

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read TELLER claims that a non-intrusive trace-and-log pipeline can pinpoint the abnormal step, the suspicious operator, and a natural-language explanation for LLM inference failures across the engine, CUDA runtime, GPU kernel, and communication

desk verdict A genuinely useful RCA architecture for LLM inference with a load-bearing evaluation leak: TPE is learned before the train/test split, so the headline F1 numbers are optimistic and the transferability claim is not yet established. read the letter →

arxiv 2608.01975 v1 pith:GRJEIFO4 submitted 2026-08-03 cs.SE cs.CLcs.LGcs.PF

Ruilin Xu , Junyi Li , Pengfei Chen , Zongxuan Xie This is my paper
classification cs.SEcs.CLcs.LGcs.PF
keywords LLMinferenceroot-causeanalysistrace-logalignmentNVTX/CUPTItracingTracePairEncodingoperatorlocalizationmultimodaldiagnosisnon-intrusiveobservability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

TELLER is a framework for root-cause analysis of LLM inference services that combines NVTX/CUPTI traces with system logs to reconstruct request-level call-chain trees, then uses a structure-preserving tokenizer (TPE) and a multimodal model to flag abnormal execution steps, localize suspicious operators, and generate explanations. The paper claims this works without modifying model binaries or engine internals, making it a practical triage substrate for engineers. If correct, it would give operators a unified view across the vertical (within-node) and horizontal (cross-node) execution stack, replacing fragmented profiler and log tools. A central finding is a compression–accuracy trade-off: moderate TPE compression reduces trace length by over 80% while preserving diagnosis quality, whereas aggressive compression severely degrades operator localization.

What carries the argument

Trace Pair Encoding (TPE): a structure-preserving tokenizer that compresses a depth-first serialization of a trace slice by merging frequent adjacent token pairs while inheriting the left token's parent index and depth, and accumulating durations. This converts an execution tree with timing into a compact token sequence with side channels (parent_idx, duration, depth) that a graph-aware encoder can consume without losing call structure.

What would settle it

Train TELLER on requests from several fault-injection runs and test on an entirely new run with a different fault timing or a different workload. If step-level F1 on the horizontal view drops from the reported 0.916 toward the classical baselines (below 0.8), the claim of practical cross-layer RCA generalization would be refuted.

Watch

Extended reading notes

Core claim

The central discovery is that cross-layer root-cause analysis for LLM inference can be structured as a request-level diagnosis task on a compact, dependency-aware representation of traces and logs. TELLER reconstructs per-request call-chain trees, aligns logs by request identifier or time window, extracts a causal-context slice preserving parent–child, temporal, and communication relations, and encodes it with TPE tokens that keep parent index, duration, and depth. A dual-head multimodal model then jointly predicts the abnormal step, localizes suspicious operators, and generates an explanation. The paper reports best-in-class step-level accuracy (0.93 horizontal, 0.91 vertical) and operator-

Load-bearing premise

The evaluation assumes that request-disjoint splits capture generalization, even though traces from the same workload, run, or fault-injection campaign can appear in both training and test sets, making the reported accuracy an in-distribution estimate rather than a proof of cross-run transferability.

Editorial extensions

If this is right

  • If TELLER's claims hold, operators can triage LLM inference incidents from non-intrusive traces and logs alone, without adding binary instrumentation or engine-specific hooks.
  • The compression–accuracy trade-off implies that trace-tokenization systems should tune vocabulary size to preserve structural distinctions, not minimize length.
  • The request-level alignment between NVTX ranges, CUPTI correlation IDs, and log timestamps offers a reusable recipe for unifying vertical and horizontal observability in other inference stacks.
  • The dual-head multimodal design suggests that fault classification and explanation generation can benefit from separate but fused trace and log streams, rather than a single flattened input.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The TPE merge rule could generalize beyond LLM inference to any hierarchical performance trace (e.g., microservice call graphs, data pipelines) where structure and timing must survive compression.
  • The paper's claim of engine-agnostic behavior is only demonstrated on a two-node A40 cluster with Qwen-based stacks; extending to larger clusters or different GPU families might reveal representation gaps that the current evaluation does not cover.
  • Because the splits are request-disjoint but not run- or workload-disjoint, the reported numbers likely overestimate generalization to unseen fault-injection campaigns; leave-one-run-out evaluation would be a stronger test.
  • The explanation audit is a reference-quality check, not an independent blinded assessment of TELLER's generated explanations, so the practical actionability in real incidents remains an open question.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. This paper presents TELLER, a pipeline for root-cause analysis of LLM inference requests. It collects NVTX/CUPTI traces and service logs without modifying model binaries, reconstructs per-request call-chain trees, aligns log lines to execution steps, extracts dependency-aware causal-context slices, and compresses them with a structure-preserving BPE-style tokenizer (TPE). A multimodal model jointly predicts abnormal steps, localizes suspicious operators, and generates natural-language explanations. The evaluation, on 300 requests / 2,482 steps across five vLLM/SGLang/Torch-FSDP settings, compares against classical and log-based baselines on horizontal and vertical views, studies TPE vocabulary size, modality ablations, low-fault priors, and overhead. The paper reports strong step-level and operator-level results at the least aggressive compression setting and positions TELLER as a triage substrate.

Significance. If taken at face value, TELLER would be a useful and well-engineered contribution: the collection design is non-intrusive in the stated sense, the causal-context slice and TPE are novel representation ideas, and the pipeline cleanly separates deterministic reconstruction from learned reasoning. The code/data availability and explicit treatment of overhead are strengths. The candidate-localization GMM feeding features into the multimodal model is a feature-engineering choice rather than circularity. However, the evaluation is currently an in-distribution study: TPE is learned before the split, the split is not run- or engine-disjoint, and the explanation audit does not validate model-generated explanations. These issues directly affect the transferability claim (R4) and the numeric comparisons, so the central claim is not yet established beyond the specific dataset.

major comments (4)
  1. [§4.2 / §3.5] Dataset construction applies TPE before the split: after expanding traces into step records, the pipeline 'applies TPE' and only then 'randomly shuffle request traces ... allocate 80%, 10%, and 10%'. Since TPE (Algorithm 2) 'repeatedly merges the most frequent adjacent pair until a target vocabulary size is reached', the merge rules and vocabulary are learned from all 300 request traces, including test requests. Test-trace motifs can therefore become token IDs that the model has seen during training, a transductive leakage that can inflate every TPE-dependent number in Tables 2, 3, 5, and 7, as well as the compression–accuracy conclusion of RQ2. The fix is straightforward: train TPE only on the training split (or on an independent trace corpus) and re-run the main comparisons.
  2. [§4.2 / §7] The paper concedes that 'correlated traces from the same workload, run, or fault-injection campaign can still cross splits, so the reported numbers should be interpreted as in-distribution estimates.' The split is request-disjoint, not run- or injection-disjoint. Because the dataset has only five settings and uses a random request split, every engine appears in both training and test; Table 7 therefore reports test performance on stacks whose traces were seen during training, not cross-engine generalization. The same issue affects Tables 2 and 3 if multiple requests from one fault-injection campaign are split across train and test. An engine-disjoint or campaign-disjoint split (plus retraining TPE on the training split) is needed to support R4.
  3. [§5.4 / §7] Explanation quality is not evaluated on TELLER-generated explanations. Table 9 checks reference explanations, not model outputs; BLEU measures lexical overlap only. The paper explicitly states the audit 'is not independent, blinded expert validation of all TELLER-generated explanations.' Since the output triplet (ŷ_s, r_s, Ô_s) includes a natural-language root-cause explanation r_s, the explanation component is a central contribution. I ask for a human evaluation of a sample of generated explanations with inter-annotator agreement, or for the paper to be re-scoped as not evaluating explanations.
  4. [§5 / Tables 2,3,5,7] The main tables report point estimates from a single run on 300 requests / 2,482 steps. No confidence intervals or repeated-seed results are given for the main comparisons. Given the acknowledged correlation structure, this makes it impossible to assess whether TELLER's margins are stable. Table 8 repeats 50 times; similar repeated runs should be reported for the main tables to support the claimed margins over baselines.
minor comments (4)
  1. [§5.1 / Table 2] The claim that TELLER achieves 'the best step-level Accuracy, Precision, and F1' is true for the horizontal view, but in the vertical view Logs2Graphs has higher Accuracy (0.922 vs 0.911) and RandomForest higher Precision (0.942 vs 0.878). The sentence should be qualified.
  2. [§4.4] The definition of Macro metrics is internally inconsistent: precision/recall for empty-empty steps are 'conservatively set to 0', yet f_i = j_i = 1 for the same case. Please clarify the exact computation, especially how F1 is defined when precision/recall are set to 0.
  3. [§3.4] The relevance weights α, β, γ, η, threshold κ, temporal scale τ₀, and the least-confident quantile q are not given values or a sensitivity analysis. As free parameters, they should be documented or shown to be stable.
  4. [§4.3 / Table 4] The 'Prompt-only' baseline is described only as removing the numeric candidate-filtering stage and asking an LLM to reason from structured summaries; please specify the prompt content, the backbone LLM, and the number of runs, since this baseline is used to motivate the two-stage design.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the diagnosis pipeline is self-contained; the noted TPE-before-split and correlated-trace issues are evaluation-validity threats, not derivation circularity.

full rationale

TELLER's derivation chain is not circular. The predicted outputs (abnormal step y_s, suspicious-operator set O_s, and explanation r_s) are supervised by explicit labels in the dataset, and the model is trained on request-disjoint splits. The TPE tokenizer and GMM candidate localizer are unsupervised preprocessing stages: TPE learns merge rules from symbol co-occurrence, and GMM computes anomaly scores from timing/count features. Neither stage uses the downstream labels, so the final predictions are not restatements of the fitted components. The paper itself discloses the main evaluation limitation: 'correlated traces from the same workload, run, or fault-injection campaign can still cross splits, so the reported numbers should be interpreted as in-distribution estimates' (Section 4.2, repeated in Section 7). Additionally, Section 4.2 applies TPE before the random request-level split, so the tokenizer vocabulary and merge rules are learned from all 300 request traces including test traces. This is a real representation-leakage threat that can inflate the reported accuracy and the compression-accuracy trade-off, but it is a benchmark-validity issue rather than a circular derivation: the test labels are not used to fit TPE or GMM, and the model's outputs are not equal to the tokenizer's or GMM's outputs by construction. There is also no load-bearing self-citation: the same-author reference [38] is not used to justify the central claims, and no uniqueness theorem or ansatz is imported from prior work. The diagnostic architecture, objective functions, and metrics are all specified internally and independently benchmarked against external baselines, so the central RCA claim has independent content.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The framework leans on several unstated modeling choices: how the GMM is fit, how relevance weights are set, and the assumption that injected faults resemble production faults. None of these are derived; they are adopted to make the pipeline work.

free parameters (5)
  • Relevance weights (alpha, beta, gamma, eta, kappa, h, tau0) in causal-context slice = not reported
    Used in Eq. (1) to compute node relevance and to prune the subgraph; no principled selection is stated.
  • GMM components per family = not reported
    Gaussian mixture model used for anomaly scoring; the number of components is not specified.
  • TPE vocabulary size = 256 (best), 512, 1024
    Vocabulary size is a hyperparameter that is varied and the best is selected; the paper concludes 256 is optimal.
  • Log alignment tolerance window = small (not quantified)
    Used to align logs to step intervals in Section 3.3; the exact tolerance is not given.
  • Least-confident quantile q in candidate family selection = not reported
    Used to select candidate family set from GMM scores; value not reported.
assumptions (4)
  • domain assumption NVTX ranges reconstruct a valid hierarchy by stack scanning
    Section 3.3 assumes push/pop events are well-formed and nested, which may not hold under asynchronous or overlapping ranges.
  • domain assumption CUPTI correlation IDs reliably bind host launches to device kernels
    Section 3.3 uses (tid, cid) matching; this assumes the correlation IDs are correct and complete.
  • domain assumption Per-request projection assigns shared operators correctly
    Section 3.3 projects descendants to requests sharing a step, which is ambiguous in many-to-one mappings.
  • domain assumption Injected faults are representative of production incidents
    Section 7 acknowledges real incidents may be longer, noisier, and more entangled than the injected scenarios.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TELLER: Non-intrusive Cross-Layer Root-Cause Analysis for LLM Inference." pith.science (2026). https://pith.science/paper/GRJEIFO4

@misc{pith2026260801975,
  author       = {Pith},
  title        = {Pith review of: TELLER: Non-intrusive Cross-Layer Root-Cause Analysis for LLM Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GRJEIFO4}},
  note         = {Machine review of arXiv:2608.01975}
}
read the original abstract

Large language model (LLM) inference has evolved from an offline workload into a continuously operated software service, yet root-cause analysis remains difficult because a single request spans the inference engine, Python/C++ backend, host CUDA APIs, GPU kernels, and distributed communication. Existing profilers expose raw timelines, while log-based diagnosis often misses cross-layer execution semantics and request-level structure. We present TELLER, a non-intrusive Trace- and Log-aware LLM inference Root-cause analysis framework. TELLER first collects NVTX/CUPTI traces and service logs without modifying model binaries, then reconstructs per-request call-chain trees and aligns log lines with the corresponding execution steps. We introduce a dependency-aware causal-context slice that preserves parent-child structure, temporal order, and communication relations, and a Trace Pair Encoding (TPE) tokenizer that compresses such slices into compact structural token sequences with parent, depth, and duration attributes. On top of these representations, TELLER combines numeric candidate localization with a multimodal root-cause model that jointly predicts abnormal steps, localizes suspicious operators, and generates natural-language explanations. Experiments on multi-node GPU inference workloads show a clear compression-accuracy trade-off: a moderate TPE vocabulary reduces per-step trace length by more than 80% while achieving the best overall performance on both horizontal (cross-node communication) and vertical (within-node execution stack) views, whereas more aggressive compression substantially degrades diagnosis quality. Further analyses under low-fault priors, strengthened baselines, modality ablations, explanation-quality checks, and tracing overhead show that TELLER provides a practical triage and evidence-localization substrate for LLM inference RCA.

Figures

Figures reproduced from arXiv: 2608.01975 by the authors.

Figure 1
Figure 1. Overview of TELLER. TELLER first performs non-intrusive cross-layer tracing of LLM inference, then reconstructs request-level call chains with aligned logs, extracts compact suspicious causal contexts, and finally conducts multimodal fault diagnosis and root-cause analysis to predict error types and generate root-cause explanations. induces a hierarchical forest that represents containment among engine steps, front-… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 2 canonical work pages

  1. [1]

    Toufique Ahmed, Supriyo Ghosh, Chetan Bansal, Thomas Zimmermann, Xuchao Zhang, and Saravan Rajmohan. 2023. Recommending Root-Cause and Mitigation Steps for Cloud Incidents Using Large Language Models. InProceedings of the 2023 IEEE/ACM 45th International Conference on Software Engineering (ICSE). IEEE, 1737–1749. doi:10.1109/ICSE48619.2023.00149

    arXiv 2023

  2. [2]

    AI21 Labs. 2026. AI21Labs. https://www.ai21.com/. Accessed: Jan. 10, 2026

    2026

  3. [3]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen Technical Report.arXiv preprint arXiv:2309.16609(2023). https://arxiv.org/abs/2309.16609

    arXiv 2023

  4. [4]

    Leo Breiman. 2001. Random Forests.Machine Learning45 (2001), 5–32. doi:10. 1023/A:1010933404324

    2001

  5. [5]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gr...

    2020

  6. [6]

    Chen, Emre Kiciman, Eugene Fratkin, Armando Fox, and Eric A

    Mike Y. Chen, Emre Kiciman, Eugene Fratkin, Armando Fox, and Eric A. Brewer

  7. [7]

    Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A Scalable Tree Boosting System. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD). ACM, 785–794. doi:10.1145/2939672. 2939785

    doi:10.1145/2939672 2016

  8. [8]

    Yinfang Chen, Huaibing Xie, Minghua Ma, Yu Kang, Xin Gao, Liu Shi, Yunjie Cao, Xuedong Gao, Hao Fan, Ming Wen, et al. 2024. Automatic Root Cause Analysis via Large Language Models for Cloud Incidents. InProceedings of the 19th European Conference on Computer Systems (EuroSys). 674–688. doi:10.1145/3627703.3629553

    arXiv 2024

Show all 44 references
  1. [9]

    Fu, Stefano Ermon, Atri Rudra, and Christopher Ré

    Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness. arXiv preprint arXiv:2205.14135(2022). doi:10.52202/068431-1189

    2022 arXiv

  2. [10]

    Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. 1996. A Density- Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise. InProceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD). 226–231. https://w...

    1996

  3. [11]

    Matthias Fey and Jan E. Lenssen. 2019. Fast Graph Representation Learning with PyTorch Geometric.arXiv preprint arXiv:1903.02428(2019). https://arxiv.org/ abs/1903.02428

    2019 arXiv

  4. [12]

    J. A. Hartigan and M. A. Wong. 1979. Algorithm AS 136: A K-Means Clustering Algorithm.Journal of the Royal Statistical Society. Series C (Applied Statistics)28, 1 (1979), 100–108. doi:10.2307/2346830

    1979 doi

  5. [13]

    Pengxiang Jin, Shenglin Zhang, Minghua Ma, Haozhe Li, Yu Kang, Liqun Li, Yudong Liu, Bo Qiao, Chaoyun Zhang, Pu Zhao, et al. 2023. Assess and Summarize: Improve Outage Understanding with Large Language Models. InProceedings of the 31st ACM Joint European Software Engineering C...

    2023

  6. [14]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. InProceedings of the 5th International Conference on Learning Representations (ICLR). https://openreview.net/forum?id=SJU4ayYgl

    2017

  7. [15]

    Taku Kudo and John Richardson. 2018. SentencePiece: A Simple and Language Independent Subword Tokenizer and Detokenizer for Neural Text Processing. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. Association for ...

    2018 doi

  8. [16]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient Memory Management for Large Language Model Serving with PagedAttention. arXiv preprint arXiv:2309.06180(2023). doi:10.1145/3600006.3613165

    2023 arXiv

  9. [17]

    Yukyung Lee, Jina Kim, and Pilsung Kang. 2023. LAnoBERT: System Log Anomaly Detection based on BERT Masked Language Model.Applied Soft Computing146 (2023), 110689. doi:10.1016/j.asoc.2023.110689

    2023

  10. [18]

    Dan Li, Dacheng Chen, Lei Shi, Baihong Jin, Jonathan Goh, and See-Kiong Ng

  11. [19]

    Zhong Li, Jiayang Shi, and Matthijs van Leeuwen. 2024. Graph Neural Net- works Based Log Anomaly Detection and Explanation. InCompanion Proceedings of the 46th IEEE/ACM International Conference on Software Engineering (ICSE- Companion). ACM, 306–307. doi:10.1145/3639478.3643084

    2024

  12. [20]

    Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. 2008. Isolation Forest. In Proceedings of the 8th IEEE International Conference on Data Mining (ICDM). IEEE, 413–422. doi:10.1109/ICDM.2008.17

    2008 doi

  13. [21]

    NVIDIA. 2024. NVIDIA CUDA Toolkit Documentation. https://docs.nvidia.com/ cuda/. Accessed: 2025-09-17

    2024

  14. [22]

    NVIDIA. 2024. NVIDIA CUPTI Documentation. https://docs.nvidia.com/cupti/ index.html. Accessed: 2025-09-17

    2024

  15. [23]

    NVIDIA. 2024. NVIDIA NVTX Documentation and Source. https://github.com/ NVIDIA/NVTX. Accessed: 2025-09-17

    2024

  16. [24]

    NVIDIA. 2025. NVIDIA Nsight Systems. https://developer.nvidia.com/nsight- systems. Accessed: Jan. 10, 2025

    2025

  17. [25]

    OpenTelemetry. 2024. OpenTelemetry: High-Quality, Ubiquitous, and Portable Telemetry to Enable Effective Observability. https://opentelemetry.io/. Accessed: Mar. 1, 2024

    2024

  18. [26]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: a Method for Automatic Evaluation of Machine Translation. InProceedings of the 40th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics, Philadelphia...

    2002

  19. [27]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Des- maison, Andreas Köpf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, L...

    2019 arXiv

  20. [28]

    PyTorch. 2025. PyTorch Profiler Recipe. https://pytorch.org/tutorials/recipes/ recipes/profiler_recipe.html. Accessed: Jan. 10, 2025

    2025

  21. [29]

    Reynolds

    Douglas A. Reynolds. 2009. Gaussian Mixture Models. InEncyclopedia of Biometrics, Stan Z. Li and Anil Jain (Eds.). Springer, Boston, MA, 659–663. doi:10.1007/978-0-387-73003-5_196

    2009 doi

  22. [30]

    Devjeet Roy, Xuchao Zhang, Rashi Bhave, Chetan Bansal, Pedro Las-Casas, Ro- drigo Fonseca, and Saravan Rajmohan. 2024. Exploring LLM-Based Agents for Root Cause Analysis. InCompanion Proceedings of the 32nd ACM Interna- tional Conference on the Foundations of Software Engineer...

    2024

  23. [31]

    Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural Machine Translation of Rare Words with Subword Units. InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Be...

    2016

  24. [32]

    Sigelman, Luiz André Barroso, Michael Burrows, Patrick Stephen- son, Manoj Plakal, Donald Beaver, Saul Jaspan, and Chandan Kumar Shanbhag

    Benjamin H. Sigelman, Luiz André Barroso, Michael Burrows, Patrick Stephen- son, Manoj Plakal, Donald Beaver, Saul Jaspan, and Chandan Kumar Shanbhag

  25. [33]

    2008.Support Vector Machines

    Ingo Steinwart and Andreas Christmann. 2008.Support Vector Machines. Springer, New York, NY. doi:10.1007/978-0-387-77242-4

    2008 doi

  26. [34]

    TensorFlow. 2020. Introducing New TensorFlow Profiler. https://blog.tensorflow. org/2020/04/introducing-new-tensorflow-profiler.html. Accessed: Jan. 10, 2025

    2020

  27. [35]

    2023.LLaMA: Open and Efficient Foundation Language Models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lam- ple. 2023.LLaMA: Open and Efficient Foundation L...

    2023 arXiv

  28. [36]

    Haopei Wang, Anubhavnidhi Abhashkumar, Changyu Lin, Tianrong Zhang, Xiaoming Gu, Ning Ma, Chang Wu, Songlin Liu, Wei Zhou, Yongbin Dong, et al

  29. [37]

    Zongxuan Xie, Junyi Li, Ruilin Xu, and Pengfei Chen. 2026. The TELLER Dataset. Dataset and code artifact. doi:10.6084/m9.figshare.33142493.v1

    2026 doi

  30. [38]

    Ruilin Xu, Zongxuan Xie, and Pengfei Chen. 2025. eACGM: Non-Instrumented Performance Tracing and Anomaly Detection Towards Machine Learning Sys- tems. In2025 IEEE/ACM 33rd International Symposium on Quality of Service (IWQoS). IEEE, 1–6. doi:10.1109/IWQoS65803.2025.11143277

    2025

  31. [39]

    Xu Zhang, Yong Xu, Qingwei Lin, Bo Qiao, Hongyu Zhang, Yingnong Dang, Chunyu Xie, Xinsheng Yang, Qian Cheng, Ze Li, Junjie Chen, Xiaoting He, Ran- dolph Yao, Jian-Guang Lou, Murali Chintalapati, Shen Furao, and Dongmei Zhang

  32. [44]

    InProceedings of the 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering (ESEC/FSE)

    Robust Log-Based Anomaly Detection on Unstable Log Data. InProceedings of the 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering (ESEC/FSE). ACM, 807–817. doi:10.1145/3338906.3338931 Received 2026-03-26; ...

    2026

  33. [2002]

    InProceedings of the 2002 International Conference on Dependable Systems and Networks (DSN)

    Pinpoint: Problem Determination in Large, Dynamic Internet Services. InProceedings of the 2002 International Conference on Dependable Systems and Networks (DSN). 595–604. doi:10.1109/DSN.2002.1029005

    2002 arXiv

  34. [2010]

    InPro- ceedings of the ACM/IEEE International Conference on Operating Systems Design and Implementation (OSDI)

    Dapper, a Large-Scale Distributed Systems Tracing Infrastructure. InPro- ceedings of the ACM/IEEE International Conference on Operating Systems Design and Implementation (OSDI). https://research.google/pubs/dapper-a-large-scale- distributed-systems-tracing-infrastructure/

  35. [2019]

    InInternational Conference on Artificial Neural Networks (ICANN)

    MAD-GAN: Multivariate Anomaly Detection for Time Series Data with Generative Adversarial Networks. InInternational Conference on Artificial Neural Networks (ICANN). Springer, 703–716. doi:10.1007/978-3-030-30490-4_56

    doi

  36. [2024]

    InProceedings of the 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI)

    NetAssistant: Dialogue Based Network Diagnosis in Data Center Networks. InProceedings of the 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI). 2011–2024. https://www.usenix.org/conference/nsdi24/ presentation/wang-haopei

    2011

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.