Pith. sign in

REVIEW 4 major objections 6 minor 76 references

Formal and Empirical Study of Metadata-Based Profiling for Resource Management in the Computing Continuum

T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper formalizes and tests a metadata-only profiler that assigns runtime profiles to workloads at submission time.

desk verdict A careful extension of the authors' own PolarisProfiler line—real but incremental—whose headline 93% accuracy is a same-trace fit, not demonstrated real-time prediction, until baselines and a temporal split are added. read the letter →

arxiv 2504.20740 v1 pith:TY2RHLIR submitted 2025-04-29 cs.DC

Andrea Morichetta , Stefan Nastic , Victor Casamayor Pujol , Schahram Dustdar This is my paper · ORCID
classification cs.DC
keywords workloadprofilingstaticmetadataresourcemanagementcomputingcontinuumHDBSCANclusteringXGBoostclassificationAlibabaPAItracesGoogleclusterdata
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

This paper claims that workload runtime behavior can be estimated in real time from static, a priori metadata alone — user, job name, workload type, task type, and group — without sandbox profiling or runtime observation. The proposed PolarisProfiler first clusters historical execution traces by runtime similarity, then learns a mapping from metadata to the resulting profile groups, and finally uses each group's quantile statistics to predict resource usage and duration for a new submission. The empirical case studies on Alibaba machine-learning traces and Google cluster data support the claim: on 10,000 unseen Alibaba workloads, the classifier assigns profiles with 95.19% accuracy and the combined runtime estimates stay below 50% error for about 93% of workloads. If correct, this gives infrastructure orchestrators in the computing continuum a fast, non-invasive way to make scheduling and provisioning decisions while still meeting SLOs.

What carries the argument

The central machinery is a two-stage model with a feedback loop. A Profile Generator clusters completed workloads by runtime features using density-based clustering (HDBSCAN on Alibaba, DBSCAN with autoencoder projection on Google), producing profile groups with per-feature statistical summaries such as percentiles and means. A Profile Classifier, implemented as XGBoost, maps static metadata features to profile group labels. A Feedback Loop monitors violations, outliers, and profile freshness, and triggers re-clustering when quality degrades. The formalization represents each workload as $w_i = (r(w_i), m(w_i))$, defines the clustering objective as minimizing intra-cluster distance, the classifier as a function $f: M \to C$, and introduces the named quality metric ACQUIRES (Algorithm's Cluster QualIty-REcall Score), which combines outlier reduction, cluster-count correctness, and the mean silhouette score.

What would settle it

Compute the same RMSE prediction test on workloads submitted after the training trace (a later month or a different cluster); if the share of workloads with error below 50% falls substantially below the reported 93%, the stationarity assumption underlying the real-time accuracy claim fails.

Watch

Extended reading notes

Core claim

The central claim is that a workload's runtime behavior can be profiled before it executes, using only metadata that is available when the workload is submitted. PolarisProfiler groups completed workloads into profile groups based on runtime similarity, trains a classifier on the static metadata of the workloads inside each group, and then assigns every new workload to a profile group from its metadata alone. The profile group provides quantile-based estimates of CPU usage, GPU usage, memory usage, and duration. On the Alibaba PAI trace, the resulting predictions have normalized RMSE below 50% for roughly 93% of 10,000 unseen workloads, and the same architecture produces strong classification and prediction results on Google cluster data, supporting the claim that the approach generalizes beyond a single dataset or workload type.

Load-bearing premise

The historical runtime traces used to build profile groups faithfully represent the runtime behavior of the workloads that will arrive later.

Editorial extensions

If this is right

  • An orchestrator can assign a newly submitted workload a runtime profile at submission time, without sandbox probing or waiting for that specific job's execution history.
  • Profiles carry quantile summaries, so resource provisioning can be based on expected variance rather than a single point estimate.
  • The feedback loop gives the system a mechanism to stay representative: when violations or outliers exceed thresholds, re-clustering refreshes the profile groups.
  • Because the formalization leaves the clusterer and classifier open, the same framework can be instantiated with different algorithms on different infrastructure traces.
  • The Google cluster result suggests the approach is not specific to machine-learning workloads, extending to general batch jobs with latency classes and resource requests.

Reading between the lines

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

  • Editorial inference: the metadata-to-profile mapping could also support admission control and overcommitment decisions, since it yields not just a point estimate but a distribution of expected resource usage.
  • Editorial inference: a stronger test would apply the trained classifier to workloads from a later period or a different data center; the paper's 10,000 test workloads come from the same two-month Alibaba trace, so distribution shift is untested.
  • Editorial inference: the SHAP analysis identifies specific users, groups, and parameter-server tasks as decision drivers, suggesting that user and group identity encode recurring job templates; this could be exploited for better estimates or could become a robustness liability if users change over time.
  • Editorial inference: the 5th-quantile prediction rule is a simple summary; the paper's own skewness analysis implies that per-profile distribution-aware estimators could reduce the high-error tail.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes PolarisProfiler, a metadata-based workload profiling approach for computing-continuum resource management. The approach first clusters historical workloads using runtime telemetry (HDBSCAN on the Alibaba case study, DBSCAN with autoencoder projection on Google data), then trains an XGBoost classifier that maps a priori static metadata (user, job name, task name, group, workload type) to the resulting profile groups. For a new workload, the assigned profile's 5th-percentile statistics are used as predictions of CPU, GPU, memory usage, and duration. The central empirical claim is that this yields an overall normalized RMSE below 50% for about 93% of classified workloads on the Alibaba dataset, with comparable results on Google cluster data. The paper also introduces the ACQUIRES clustering-quality metric and a feedback-loop mechanism, and it makes the implementation publicly available.

Significance. If the empirical claim were established, the approach would be a useful contribution: it is non-invasive, operates in real time, uses only metadata available at submission time, and is evaluated on two large public traces with released code. The formalization of the pipeline components provides a clear vocabulary for reasoning about profile generation, classification, and feedback. The extension to Google cluster data with different clustering mechanisms supports the claim of general applicability. However, the significance is conditional on the evaluation convincingly demonstrating out-of-sample generalization of the prediction, which the current design and reporting do not establish.

major comments (4)
  1. [Section 3.5; Abstract] The headline claim that 'our methodology yields an overall error rate below 50% for the 93% of classified workloads' is presented as evidence of accurate real-time estimation for newly submitted workloads. However, the 10,000 'unseen' test workloads are drawn from the same two-month Alibaba trace used for clustering (Section 3.3) and classifier training (Section 3.4); there is no temporal or distributional split. A classifier and profile set built on that trace cannot be distinguished by this experiment from a system that merely memorizes the trace's empirical distribution. Section 6 lists long-time evolution as future work, confirming that distribution shift is untested. To support the real-time claim, the evaluation should include a temporal split (e.g., train on an earlier window and test on a later window) or an explicit demonstration of robustness to distribution shift.
  2. [Section 3.5] The prediction summary statistic is selected on the test set: 'After experimenting with different settings, the approach that gave us the best results is to use the 5th quantile for the prediction.' This is post-hoc model selection on the evaluation data and inflates the reported accuracy. In addition, no trivial baseline is reported: a global 5th percentile of historical feature values, a per-workload-type median, or even a constant prediction would likely also yield a large fraction of workloads below the 50% normalized-RMSE threshold for heavy-tailed resource-usage distributions. Without such baselines, the 93% figure does not demonstrate that the metadata-based profile assignment improves prediction. Please report baselines and select the quantile on a validation split.
  3. [Section 3.5; Figures 14-15] The definition of the 'overall' normalized RMSE is missing. The text names RMSE_perc and reports CDFs and box plots for the four features 'altogether,' but no formula is given for how the four per-feature normalized errors are aggregated into a single value, nor is the normalization for each feature defined. This makes the headline 93% value non-reproducible. Moreover, the box plots in Figures 14 and 15 show only 16 profile labels (0,1,3,7,8,9,10,12,13,14,16,19,20,21,23,25), whereas Table 3 reports classification for 26 profiles; the paper should state how the 10,000 test workloads are distributed across profiles and how the denominator for the 93% claim is computed, especially given the exclusion of the outliers group and of profiles with near-zero F1-scores in Table 3.
  4. [Section 2.3.1; Section 3.3.2] The ACQUIRES metric is used to select HDBSCAN over OPTICS in the Profile Generator grid search, but it introduces a free parameter C_optimal (the 'ideal' number of clusters) and equal component weights that are not justified. For HDBSCAN, which does not take a cluster count as input, the paper does not specify how C_optimal is obtained, and no sensitivity analysis is provided for the equal-weight assumption. Since the choice of clustering algorithm and hyperparameters shapes the profile groups and hence the downstream classifier, this is a load-bearing step of the evaluation; the authors should either justify C_optimal and the weights, or show that the final prediction results are robust to their variation.
minor comments (6)
  1. [Section 3.5] The text says 'Figure?? represents the ECDF' and Section 2.3 contains 'Figure ?? highlights this perspective'; these missing cross-references should be fixed.
  2. [Section 2.1] The phrase 'groups together various types of workload workload' contains a duplicated word and should be corrected.
  3. [Section 3.5] The sentence listing the predicted features says 'We do so for duration, CPU usage, and GPU usage' but memory usage is also reported in Figures 13c and 14c; clarify whether memory is predicted with the same 5th-quantile rule.
  4. [Section 4] The Google dataset results are described as 'similar performance' but no aggregate percentage of workloads below the 50% RMSE threshold is reported; report the same summary statistic for comparability.
  5. [Section 2.3] The freshness condition in the UT(t) equation uses 'min(CS(C_i,t)) < tau_f' but CS is not defined; define it or use the previously defined F(C_i,t).
  6. [Section 5.3] The phrase 'Height-weight-weight' appears twice and is likely a typo for a computational graph description.

Circularity Check

0 steps flagged · score 2.0 of 10

No constructional circularity: the profiling pipeline is trained on historical runtime traces and evaluated on held-out workloads; the main caveats are a same-trace split and test-set selection of the 5th-quantile rule, which limit generalization claims but do not make the prediction equivalent to its inputs.

full rationale

The paper's derivation chain is: (1) cluster workloads by runtime features (Section 3.3), (2) train an XGBoost classifier to map static metadata to cluster labels (Section 3.4), and (3) for a test workload, use the assigned profile's historical 5th percentile as the predicted runtime value and compare it with the workload's actual value (Section 3.5). Each stage uses inputs distinct from the quantity being predicted: the classifier is trained on metadata labels derived from runtime clustering, and the predicted quantiles are computed from the training profiles, not from the test workloads' actual values. The evaluation uses 10,000 'unseen' workloads (Section 3.5), so the headline 93% figure is not forced by construction. The main validity limitations are non-circular: the test workloads are drawn from the same two-month Alibaba trace used to build the profiles, with no temporal distribution shift, and the prediction rule was selected post hoc on the test set ('After experimenting with different settings, the approach that gave us the best results is to use the 5th quantile for the prediction'), so the reported accuracy may be optimistically biased and is not demonstrated for future workloads; Section 6 acknowledges long-time evolution as future work. The paper also builds on the authors' earlier PolarisProfiler paper [44], but this self-citation is not load-bearing: the current work presents the model in full and adds new empirical evaluation on a larger dataset and the Google cluster traces, so the central claim does not reduce to the cited prior work.

Assumptions & free parameters 8 free parameters · 5 assumptions · 1 invented entities

The central claim rests mostly on domain assumptions about metadata informativeness and stationarity of workload behavior, plus several tuned parameters (cluster size, transform, distance, prediction quantile, violation thresholds). No new physical entities are introduced; ACQUIRES is a new composite metric without external validation.

free parameters (8)
  • HDBSCAN minimum cluster size = 300 (Alibaba case study)
    Selected by ACQUIRES grid search over {50,100,200,300,400,600,1000}; affects cluster granularity and downstream prediction errors.
  • Data transform function = PowerTransform
    Chosen among four scalers (Standard, MinMax, Robust, Power) by ACQUIRES grid search.
  • Distance metric = Euclidean
    Chosen over Manhattan by ACQUIRES grid search for HDBSCAN.
  • Prediction summary quantile = 5th percentile (median if skewness <= 1)
    Selected after experimenting with different settings in Section 3.5; not derived from first principles.
  • Violation rate threshold (tau_v) = 0.01 (1000 violations out of 10,000)
    Chosen to trigger re-clustering in the feedback loop experiment (Section 3.6); no sensitivity analysis.
  • ACQUIRES weights = w1=w2=w3=1/3
    Default equal weighting assumption stated in Section 2.3.1, adjustable per use case.
  • Ideal cluster count C_optimal in ACQUIRES = Not specified
    The metric requires an ideal cluster count input, but the paper does not state how C_optimal is set in the evaluations.
  • Autoencoder latent dimension (Google) = 3
    Chosen to reduce runtime features before DBSCAN on Google data (Section 4); no justification.
assumptions (5)
  • domain assumption Static metadata features (job name, user, group, task name, workload) are available at workload submission and are informative of runtime behavior.
    Core premise of the Profile Classifier (Section 2.2 and 3.4); if metadata is absent or uninformative, classification degrades.
  • domain assumption Historical runtime traces are representative of future workload behavior (stationarity within the evaluation window).
    Profile Generator clusters historical data (Section 2.1), and the test set is drawn from the same trace; the paper itself flags long-time evolution as future work (Section 6).
  • domain assumption Workloads that are similar in runtime features form meaningful clusters that can be predicted from metadata.
    The entire profile-group prediction rests on this clusterability premise, supported empirically by Hopkins statistic (Section 3.1) and silhouette scores.
  • ad hoc to paper The 5th percentile of a profile is an appropriate point estimate for resource usage and duration.
    Chosen after experimentation (Section 3.5); serves as the final prediction, so the headline error rate depends on this choice.
  • standard math Clustering algorithms (HDBSCAN/OPTICS/DBSCAN), XGBoost, and SHAP behave as documented in their references.
    Standard off-the-shelf tools used without modification.
invented entities (1)
  • ACQUIRES composite clustering quality metric
    purpose: Provides a single quality score for profile-group clustering and triggers feedback-loop updates when quality degrades.
    Introduced in Section 2.3.1; combines silhouette score, outlier fraction, and cluster-count deviation with equal weights. It is not compared against existing composite clustering metrics beyond a mention of HCS, so it has no independent validation outside this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Formal and Empirical Study of Metadata-Based Profiling for Resource Management in the Computing Continuum." pith.science (2026). https://pith.science/paper/TY2RHLIR

@misc{pith2026250420740,
  author       = {Pith},
  title        = {Pith review of: Formal and Empirical Study of Metadata-Based Profiling for Resource Management in the Computing Continuum},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TY2RHLIR}},
  note         = {Machine review of arXiv:2504.20740}
}
read the original abstract

We present and formalize a general approach for profiling workload by leveraging only a priori available static metadata to supply appropriate resource needs. Understanding the requirements and characteristics of a workload's runtime is essential. Profiles are essential for the platform (or infrastructure) provider because they want to ensure that Service Level Agreements and their objectives (SLOs) are fulfilled and, at the same time, avoid allocating too many resources to the workload. When the infrastructure to manage is the computing continuum (i.e., from IoT to Edge to Cloud nodes), there is a big problem of placement and tradeoff or distribution and performance. Still, existing techniques either rely on static predictions or runtime profiling, which are proven to deliver poor performance in runtime environments or require laborious mechanisms to produce fast and reliable evaluations. We want to propose a new approach for it. Our profile combines the information from past execution traces with the related workload metadata, equipping an infrastructure orchestrator with a fast and precise association of newly submitted workloads. We differentiate from previous works because we extract the profile group metadata saliency from the groups generated by grouping similar runtime behavior. We first formalize its functioning and its main components. Subsequently, we implement and empirically analyze our proposed technique on two public data sources: Alibaba cloud machine learning workloads and Google cluster data. Despite relying on partially anonymized or obscured information, the approach provides accurate estimates of workload runtime behavior in real-time.

Figures

Figures reproduced from arXiv: 2504.20740 by the authors.

Figure 1
Figure 1. Overview of the PolarisProfiler’s model. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visual representation of the PolarisProfiler components, actors and their interactions in the model lifecycle. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Flowchart diagram of the definition of the Profile Generator. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Flowchart diagram of the definition of the Profile Classifier. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Flowchart diagram of the process of feedback loop for the PolarisProfiler. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Flowchart diagram from the perspective of the application’s workload. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: The plots depict the relationship between the main search parameters and the final score. The purple solid line and the blue [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: The silhouette coefficient scores for the points in the extracted clusters. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Box plots representing the distribution in the clustered profiles of CPU, GPU, Memory, and Duration. The y axis is in logarithmic [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Heatmaps reporting the distribution of the main values for the metadata features on the extracted profiles. Axis labels have [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: SHAP summary plot. trained model in predicting profiles independently from their size and starting just from a priori knowledge about a workload and its instance(s). This result gives us a promising path towards reproducing the proposed profiling approach, given the s…
Figure 12
Figure 12. Figure 12: Bar plots representing the skewness in the clustered profiles of CPU, GPU, Memory, and Duration. [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 13
Figure 13. Figure 13: CDFs of the 𝑅𝑀𝑆𝐸𝑝𝑒𝑟𝑐 values for the four considered measures: CPU, GPU, Memory, and Duration. Considering features in isolation. Figure?? represents the Empirical Cumulative Distribution Function (ECDF) of the 𝑅𝑀𝑆𝐸𝑝𝑒𝑟𝑐 values for the four different metrics. The x-axis…
Figure 14
Figure 14. Figure 14: Boxplots of the 𝑅𝑀𝑆𝐸𝑝𝑒𝑟𝑐 values for the four considered measures: CPU, GPU, Memory, and Duration [PITH_FULL_IMAGE:figures/full_fig_p022_14.png]
Figure 15
Figure 15. Figure 15: CDF and boxplots representing the 𝑅𝑀𝑆𝐸𝑝𝑒𝑟𝑐 values by profile, considering 𝑅𝑀𝑆𝐸𝑝𝑒𝑟𝑐 computed on the four main features alltogether. Considering the overall result for the four features combined. Here, we analyze the 𝑅𝑀𝑆𝐸𝑝𝑒𝑟𝑐 when considering all four measures in combin…
Figure 16
Figure 16. Figure 16: Plots of the 𝑅𝑀𝑆𝐸𝑝𝑒𝑟𝑐 values on the Google cluster data for CPU, Memory, and Duration. job) scheduled on a node as a set of tasks. All tasks within a job execute the same binary, sharing the same options and requests. Hence, different task categories run as separate j…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 73 canonical work pages

  1. [1]

    A., Carey, M

    Abadi, D., Agrawal, R., Ailamaki, A., Balazinska, M., Bernstein, P. A., Carey, M. J., Chaudhuri, S., Dean, J., Doan, A., Franklin, M. J., et al. The beckman report on database research. Communications of the ACM 59 , 2 (2016), 92–99

    work page 2016

  2. [2]

    Ahmad, S., Guan, H., and Sitaraman, R. K. Loki: A system for serving ml inference pipelines with hardware and accuracy scaling. In Proceedings of the 33rd International Symposium on High-Performance Parallel and Distributed Computing (2024), pp. 267–280

    work page 2024

  3. [3]

    K., and Butt, A

    Albahar, H., Dongare, S., Du, Y., Zhao, N., Paul, A. K., and Butt, A. R. Schedtune: A heterogeneity-aware gpu scheduler for deep learning. In 2022 22nd IEEE International Symposium on Cluster, Cloud and Internet Computing (CCGrid) (2022), IEEE, pp. 695–705

    work page 2022

  4. [4]

    M., Kriegel, H.-P., and Sander, J

    Ankerst, M., Breunig, M. M., Kriegel, H.-P., and Sander, J. Optics: ordering points to identify the clustering structure. In ACM Sigmod record (1999), vol. 28, ACM, pp. 49–60

    work page 1999

  5. [5]

    Banerjee, A., and Dave, R. N. Validating clusters using the hopkins statistic. In 2004 IEEE International conference on fuzzy systems (IEEE Cat. No. 04CH37542) (2004), vol. 1, IEEE, pp. 149–153

    work page 2004

  6. [6]

    Bartzas, A., Peon-Quiros, M., Poucet, C., Baloukas, C., Mamagkakis, S., Catthoor, F., Soudris, D., and Mendias, J. M. Software metadata: Systematic characterization of the memory behaviour of dynamic applications. Journal of Systems and Software 83 , 6 (2010), 1051–1075

    work page 2010

  7. [7]

    M., Gebru, T., McMillan-Major, A., and Shmitchell, S

    Bender, E. M., Gebru, T., McMillan-Major, A., and Shmitchell, S. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (2021), pp. 610–623

    work page 2021

  8. [8]

    L., Wang, C., and Youssef, A.{AI4DL}: Mining behaviors of deep learning workloads for resource management

    Berral, J. L., Wang, C., and Youssef, A.{AI4DL}: Mining behaviors of deep learning workloads for resource management. In 12th USENIX Workshop on Hot Topics in Cloud Computing (HotCloud 20) (2020)

    work page 2020

Show all 76 references
  1. [9]

    C., Massari, L., and Tessera, D

    Calzarossa, M. C., Massari, L., and Tessera, D. Workload characterization: A survey revisited. ACM Computing Surveys (CSUR) 48 , 3 (2016), 1–43

    2016

  2. [10]

    J., Moulavi, D., and Sander, J

    Campello, R. J., Moulavi, D., and Sander, J. Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining (2013), Springer, pp. 160–172

    2013

  3. [11]

    In 2023 18th Annual System of Systems Engineering Conference (SOSE), (to appear) (2023)

    Casamayor Pujol, V., Morichetta, A., and Nastic, S.Intelligent sampling: A novel approach to optimize workload scheduling in large-scale heterogeneous computing continuum. In 2023 18th Annual System of Systems Engineering Conference (SOSE), (to appear) (2023)

    2023

  4. [12]

    Chen, S., Delimitrou, C., and Martínez, J. F. Parties: Qos-aware resource partitioning for multiple interactive services. In Proceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating Systems (2019), pp. 107–120

    2019

  5. [13]

    Analysis and characterization of performance variability for openmp runtime

    Cui, M., Papadopoulou, N., and Pericàs, M. Analysis and characterization of performance variability for openmp runtime. In Proceedings of the SC’23 Workshops of The International Conference on High Performance Computing, Network, Storage, and Analysis (2023), pp. 1614–1622

    2023

  6. [14]

    Kairos: Preemptive data center scheduling without runtime estimates

    Delgado, P., Didona, D., Dinu, F., and Zwaenepoel, W. Kairos: Preemptive data center scheduling without runtime estimates. In Proceedings of the ACM Symposium on Cloud Computing (2018), pp. 135–148

    2018

  7. [15]

    Achieving horizontal scalability in density-based clustering for urls

    Faroughi, A., Javidan, R., Mellia, M., Morichetta, A., Soro, F., and Trevisan, M. Achieving horizontal scalability in density-based clustering for urls. In 2018 IEEE International Conference on Big Data (Big Data) (2018), IEEE, pp. 3841–3846

    2018

  8. [16]

    D., Bodik, P., Kandula, S., Boutin, E., and Fonseca, R

    Ferguson, A. D., Bodik, P., Kandula, S., Boutin, E., and Fonseca, R. Jockey: guaranteed job latency in data parallel clusters. In Proceedings of the 7th ACM european conference on Computer Systems (2012), pp. 99–112

    2012

  9. [17]

    Y., Gao, Y., Golikov, P., and Pekhimenko, G.Habitat: A{Runtime-Based} computational performance predictor for deep neural network training

    Geoffrey, X. Y., Gao, Y., Golikov, P., and Pekhimenko, G.Habitat: A{Runtime-Based} computational performance predictor for deep neural network training. In 2021 USENIX Annual Technical Conference (USENIX ATC 21) (2021), pp. 503–521

    2021

  10. [18]

    P., Li, M., Zhang, L., and Ardagna, D

    Gibilisco, G. P., Li, M., Zhang, L., and Ardagna, D. Stage aware performance modeling of dag based in memory analytic platforms. In 2016 IEEE 9th International Conference on Cloud Computing (CLOUD) (2016), IEEE, pp. 188–195

    2016

  11. [19]

    A profile based network intrusion detection and prevention system for securing cloud environment

    Gupta, S., Kumar, P., and Abraham, A. A profile based network intrusion detection and prevention system for securing cloud environment. International Journal of Distributed Sensor Networks 9 , 3 (2013), 364575

    2013

  12. [20]

    On cloud-based oversubscription

    Householder, R., Arnold, S., and Green, R. On cloud-based oversubscription. International Journal of Engineering Trends and Technology (IJETT) Manuscript submitted to ACM Formal and Empirical Study of Metadata-Based Profiling for Resource Management in the Computing Continuum ...

    2014

  13. [21]

    In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (2021), pp

    Hu, Q., Sun, P., Yan, S., Wen, Y., and Zhang, T.Characterization and prediction of deep learning workloads in large-scale gpu datacenters. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (2021), pp. 1–15

    2021

  14. [22]

    Detecting layered bottlenecks in microservices

    Inagaki, T., Ueda, Y., Ohara, M., Choochotkaew, S., Amaral, M., Trent, S., Chiba, T., and Zhang, Q. Detecting layered bottlenecks in microservices. In 2022 IEEE 15th International Conference on Cloud Computing (CLOUD) (2022), IEEE, pp. 385–396

    2022

  15. [23]

    C., Lin, X., and Deng, N

    Jajoo, A., Hu, Y. C., Lin, X., and Deng, N. A case for task sampling based learning for cluster job scheduling. In 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22) (2022), pp. 19–33

    2022

  16. [24]

    Network-aware scheduling for data-parallel jobs: Plan when you can

    Jalaparti, V., Bodik, P., Menache, I., Rao, S., Makarychev, K., and Caesar, M. Network-aware scheduling for data-parallel jobs: Plan when you can. ACM SIGCOMM Computer Communication Review 45 , 4 (2015), 407–420

    2015

  17. [25]

    Determining the optimal number of clusters using silhouette score as a data mining technique

    Januzaj, Y., Beqiri, E., and Luma, A. Determining the optimal number of clusters using silhouette score as a data mining technique. International Journal of Online & Biomedical Engineering 19 , 4 (2023)

    2023

  18. [26]

    A., Curino, C., Menache, I., Narayanamurthy, S

    Jyothi, S. A., Curino, C., Menache, I., Narayanamurthy, S. M., Tumanov, A., Yaniv, J., Mavlyutov, R., Goiri, I., Krishnan, S., Kulkarni, J., et al. Morpheus: Towards automated{SLOs} for enterprise clusters. In 12th USENIX Symposium on Operating Systems Design and Implementatio...

    2016

  19. [27]

    M., Heddaya, S., Ramakrishnan, R., and Sakalanaga, S

    Karanasos, K., Rao, S., Curino, C., Douglas, C., Chaliparambil, K., Fumarola, G. M., Heddaya, S., Ramakrishnan, R., and Sakalanaga, S. Mercury: Hybrid centralized and distributed scheduling in large shared clusters. In 2015 USENIX Annual Technical Conference (USENIX ATC 15) (2...

    2015

  20. [28]

    K., and Simha, A

    Kattepur, A., Rath, H. K., and Simha, A. A-priori estimation of computation times in fog networked robotics. In 2017 IEEE international conference on edge computing (EDGE) (2017), IEEE, pp. 9–16

    2017

  21. [29]

    A study of contributing factors to power aware vertical scaling of deadline constrained applications

    Kaushik, P., Raghavendra, S., and Govindaraju, M. A study of contributing factors to power aware vertical scaling of deadline constrained applications. In 2022 IEEE 15th International Conference on Cloud Computing (CLOUD) (2022), IEEE, pp. 500–510

    2022

  22. [30]

    Weblens: Towards interactive large-scale structured data profiling

    Khan, R., and Gubanov, M. Weblens: Towards interactive large-scale structured data profiling. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (2020), pp. 3425–3428

    2020

  23. [31]

    Workload forecasting and energy state estimation in cloud data centres: Ml-centric approach

    Khan, T., Tian, W., Ilager, S., and Buyya, R. Workload forecasting and energy state estimation in cloud data centres: Ml-centric approach. Future Generation Computer Systems 128 (2022), 320–332

    2022

  24. [32]

    In 20th USENIX Conference on File and Storage Technologies (FAST 22) (2022), pp

    Kisous, R., Kolikant, A., Duggal, A., Sheinvald, S., and Y adgar, G.The what, the from, and the to: The migration games in deduplicated systems. In 20th USENIX Conference on File and Storage Technologies (FAST 22) (2022), pp. 265–280

    2022

  25. [33]

    Time series predictions for cloud workloads: A comprehensive evaluation

    Lackinger, A., Morichetta, A., and Dustdar, S. Time series predictions for cloud workloads: A comprehensive evaluation. In 2024 IEEE International Conference on Service-Oriented System Engineering (SOSE) (2024), IEEE, p. 00

    2024

  26. [34]

    arXiv preprint arXiv:2202.07896 (2022)

    Li, J., Xu, H., Zhu, Y., Liu, Z., Guo, C., and W ang, C.Aryl: An elastic cluster scheduler for deep learning. arXiv preprint arXiv:2202.07896 (2022)

    2022 arXiv

  27. [35]

    D., Alistarh, D., and Hoefler, T

    Li, S., Ben-Nun, T., Girolamo, S. D., Alistarh, D., and Hoefler, T. Taming unbalanced training workloads in deep learning with partial collective operations. In Proceedings of the 25th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (2020), pp. 45–61

    2020

  28. [36]

    M., Erion, G., Chen, H., DeGrave, A., Prutkin, J

    Lundberg, S. M., Erion, G., Chen, H., DeGrave, A., Prutkin, J. M., Nair, B., Katz, R., Himmelfarb, J., Bansal, N., and Lee, S.-I. From local explanations to global understanding with explainable ai for trees. Nature machine intelligence 2 , 1 (2020), 56–67

    2020

  29. [37]

    M., and Lee, S.-I

    Lundberg, S. M., and Lee, S.-I. A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017)

    2017

  30. [38]

    Themis: Fair and efficient {GPU} cluster scheduling

    Mahajan, K., Balasubramanian, A., Singhvi, A., Venkataraman, S., Akella, A., Phanishayee, A., and Chawla, S. Themis: Fair and efficient {GPU} cluster scheduling. In 17th USENIX Symposium on Networked Systems Design and Implementation (NSDI 20) (2020), pp. 289–304

    2020

  31. [39]

    Optimizing cloud function configuration via local simulations

    Manner, J., Endre𝛽, M., Böhm, S., and Wirtz, G. Optimizing cloud function configuration via local simulations. In 2021 IEEE 14th International Conference on Cloud Computing (CLOUD) (2021), IEEE, pp. 168–178

    2021

  32. [40]

    A hybrid score to optimize clustering hyperparameters for online search term data

    Martino, A., and Rossetto, A. A hybrid score to optimize clustering hyperparameters for online search term data. In 2022 IEEE International Conference on Big Data (Big Data) (2022), IEEE, pp. 2317–2322

    2022

  33. [41]

    Accelerated hierarchical density based clustering

    McInnes, L., and Healy, J. Accelerated hierarchical density based clustering. In 2017 IEEE International Conference on Data Mining Workshops (ICDMW) (Nov 2017), pp. 33–42

    2017

  34. [42]

    Flexllm: A system for co-serving large language model inference and parameter- efficient finetuning

    Miao, X., Oliaro, G., Cheng, X., Wu, M., Unger, C., and Jia, Z. Flexllm: A system for co-serving large language model inference and parameter- efficient finetuning. arXiv preprint arXiv:2402.18789 (2024)

    2024

  35. [43]

    Lenta: Longitudinal exploration for network traffic analysis from passive data

    Morichetta, A., and Mellia, M. Lenta: Longitudinal exploration for network traffic analysis from passive data. IEEE Transactions on Network and Service Management 16, 3 (2019), 814–827

    2019

  36. [44]

    Morichetta, A., Pujol, V. C., Nastic, S., Dustdar, S., Vij, D., Xiong, Y., and Zhang, Z.Polarisprofiler: A novel metadata-based profiling approach for optimizing resource management in the edge-cloud continnum. 2023 IEEE International Conference on Service-Oriented System Engi...

    2023

  37. [45]

    C., Raith, P., Xiong, Y., Nastic, S., Dustdar, S., and Zhang, Z.Demystifying deep learning in predictive monitoring for cloud-native slos

    Morichetta, A., Pusztai, T., Vij, D., Pujol, V. C., Raith, P., Xiong, Y., Nastic, S., Dustdar, S., and Zhang, Z.Demystifying deep learning in predictive monitoring for cloud-native slos. In 2023 IEEE 16th International Conference on Cloud Computing (CLOUD) (2023), IEEE, pp. 1–11

    2023

  38. [46]

    In 14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20) (2020), pp

    Narayanan, D., Santhanam, K., Kazhamiaka, F., Phanishayee, A., and Zaharia, M.{Heterogeneity-Aware} cluster scheduling policies for deep learning workloads. In 14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20) (2020), pp. 481–498

    2020

  39. [47]

    In4th IEEE International Conference on Cognitive Machine Intelligence (CogMi) (2022)

    Nastic, S., Dustdar, S., Philipp, R., Alireza, F., and Pusztai, T.A serverless computing fabric for edge & cloud. In4th IEEE International Conference on Cognitive Machine Intelligence (CogMi) (2022). Manuscript submitted to ACM 30 Morichetta et al

    2022

  40. [48]

    Sloc: Service level objectives for next generation cloud computing

    Nastic, S., Morichetta, A., Pusztai, T., Dustdar, S., Ding, X., Vij, D., and Xiong, Y. Sloc: Service level objectives for next generation cloud computing. IEEE Internet Computing 24 , 3 (2020), 39–50

    2020

  41. [49]

    C., Dustdar, S., Vij, D., and Xiong, Y.Polaris scheduler: Edge sensitive and slo aware workload scheduling in cloud-edge-iot clusters

    Nastic, S., Pusztai, T., Morichetta, A., Pujol, V. C., Dustdar, S., Vij, D., and Xiong, Y.Polaris scheduler: Edge sensitive and slo aware workload scheduling in cloud-edge-iot clusters. In 2021 IEEE 14th International Conference on Cloud Computing (CLOUD) (2021), IEEE, pp. 206–216

    2021

  42. [50]

    Sparse autoencoder

    Ng, A., et al. Sparse autoencoder. CS294A Lecture notes 72 , 2011 (2011), 1–19

    2011

  43. [51]

    Noonan, S. M. Managing resource bursting, Aug. 16 2016. US Patent 9,417,902

    2016

  44. [52]

    In 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22) (2022), pp

    Pang, W., Panda, S., Amjad, J., Diot, C., and Govindan, R.{CloudCluster}: Unearthing the functional structure of a cloud service. In 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22) (2022), pp. 1213–1230

    2022

  45. [53]

    {DeepSketch}: A new machine {Learning-Based} reference search technique for {Post- Deduplication} delta compression

    Park, J., Kim, J., Kim, Y., Lee, S., and Mutlu, O. {DeepSketch}: A new machine {Learning-Based} reference search technique for {Post- Deduplication} delta compression. In 20th USENIX Conference on File and Storage Technologies (FAST 22) (2022), pp. 247–264

    2022

  46. [54]

    W., Tumanov, A., Jiang, A., Kozuch, M

    Park, J. W., Tumanov, A., Jiang, A., Kozuch, M. A., and Ganger, G. R. 3sigma: distribution-based cluster scheduling for runtime uncertainty. In Proceedings of the Thirteenth EuroSys Conference (2018), pp. 1–17

    2018

  47. [55]

    In IEEE 14th International Conference on Cloud Computing (CLOUD) (2021)

    Pusztai, T., Nastic, S., Morichetta, A., Casamayor Pujol, V., Dustdar, S., Ding, X., Vij, D., and Xiong, Y.A novel middleware for efficiently implementing complex cloud-native slos. In IEEE 14th International Conference on Cloud Computing (CLOUD) (2021)

    2021

  48. [56]

    Serverless edge computing—where we are and what lies ahead

    Raith, P., Nastic, S., and Dustdar, S. Serverless edge computing—where we are and what lies ahead. IEEE Internet Computing 27 , 3 (2023), 50–64

    2023

  49. [57]

    In 2021 IEEE 14th International Conference On Cloud Computing (CLOUD) (2021), IEEE, pp

    Rao, B., Liu, Z., Zhang, H., Lu, S., and W ang, L.Soda: A semantics-aware optimization framework for data-intensive applications using hybrid program analysis. In 2021 IEEE 14th International Conference On Cloud Computing (CLOUD) (2021), IEEE, pp. 433–444

    2021

  50. [58]

    A tale of two scales: Reconciling horizontal and vertical scaling for inference serving systems

    Razavi, K., Salmani, M., Mühlhäuser, M., Koldehofe, B., and Wang, L. A tale of two scales: Reconciling horizontal and vertical scaling for inference serving systems. arXiv preprint arXiv:2407.14843 (2024)

    2024 arXiv

  51. [59]

    J., and Kozyrakis, C.{INFaaS}: Automated model-less inference serving

    Romero, F., Li, Q., Yadwadkar, N. J., and Kozyrakis, C.{INFaaS}: Automated model-less inference serving. In 2021 USENIX Annual Technical Conference (USENIX ATC 21) (2021), pp. 397–411

    2021

  52. [60]

    Rousseeuw, P. J. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis.Journal of computational and applied mathematics 20 (1987), 53–65

    1987

  53. [61]

    In 2022 IEEE 15th International Conference on Cloud Computing (CLOUD) (2022), IEEE, pp

    Shin, C., Yang, G., Yoo, Y., Lee, J., and Yoo, C.Xonar: Profiling-based job orderer for distributed deep learning. In 2022 IEEE 15th International Conference on Cloud Computing (CLOUD) (2022), IEEE, pp. 112–114

    2022

  54. [62]

    F., Beskales, G., Cherniack, M., Zdonik, S

    Stonebraker, M., Bruckner, D., Ilyas, I. F., Beskales, G., Cherniack, M., Zdonik, S. B., Pagan, A., and Xu, S.Data curation at scale: the data tamer system. In Cidr (2013), vol. 2013

    2013

  55. [63]

    J., and Zhang, B.Automatic database management system tuning through large-scale machine learning

    V an Aken, D., Pavlo, A., Gordon, G. J., and Zhang, B.Automatic database management system tuning through large-scale machine learning. In Proceedings of the 2017 ACM international conference on management of data (2017), pp. 1009–1024

    2017

  56. [64]

    arXiv preprint arXiv:2401.05831 (2024)

    V ardakas, G., Pavlopoulos, J., and Likas, A.Revisiting silhouette: From micro to macro aggregation. arXiv preprint arXiv:2401.05831 (2024)

    2024 arXiv

  57. [65]

    Verbraeken, J., Wolting, M., Katzy, J., Kloppenburg, J., Verbelen, T., and Rellermeyer, J. S. A survey on distributed machine learning. Acm computing surveys (csur) 53, 2 (2020), 1–33

    2020

  58. [66]

    Large-scale cluster management at google with borg

    Verma, A., Pedrosa, L., Korupolu, M., Oppenheimer, D., Tune, E., and Wilkes, J. Large-scale cluster management at google with borg. In Proceedings of the tenth european conference on computer systems (2015), pp. 1–17

    2015

  59. [67]

    Are machine learning cloud apis used correctly? In 2021 IEEE/ACM 43rd International Conference on Software Engineering (ICSE) (2021), IEEE, pp

    Wan, C., Liu, S., Hoffmann, H., Maire, M., and Lu, S. Are machine learning cloud apis used correctly? In 2021 IEEE/ACM 43rd International Conference on Software Engineering (ICSE) (2021), IEEE, pp. 125–137

    2021

  60. [68]

    In 9th USENIX Workshop on Hot Topics in Cloud Computing (HotCloud

    W ang, H., and Li, B.Lube: Mitigating bottlenecks in wide area data analytics. In 9th USENIX Workshop on Hot Topics in Cloud Computing (HotCloud

  61. [69]

    In 2019 IEEE international symposium on workload characterization (IISWC) (2019), IEEE, pp

    W ang, M., Meng, C., Long, G., Wu, C., Y ang, J., Lin, W., and Jia, Y.Characterizing deep learning training workloads on alibaba-pai. In 2019 IEEE international symposium on workload characterization (IISWC) (2019), IEEE, pp. 189–202

    2019

  62. [70]

    In 19th{USENIX} Symposium on Networked Systems Design and Implementation ({NSDI} 22) (2022)

    Weng, Q., Xiao, W., Yu, Y., W ang, W., W ang, C., He, J., Li, Y., Zhang, L., Lin, W., and Ding, Y.MLaaS in the wild: Workload analysis and scheduling in large-scale heterogeneous GPU clusters. In 19th{USENIX} Symposium on Networked Systems Design and Implementation ({NSDI} 22) (2022)

    2022

  63. [71]

    In 13th USENIX Symposium on Operating Systems Design and Implementation (OSDI 18) (2018), pp

    Xiao, W., Bhardwaj, R., Ramjee, R., Sivathanu, M., Kwatra, N., Han, Z., Patel, P., Peng, X., Zhao, H., Zhang, Q., et al.Gandiva: Introspective cluster scheduling for deep learning. In 13th USENIX Symposium on Operating Systems Design and Implementation (OSDI 18) (2018), pp. 595–610

    2018

  64. [72]

    Extend cloud to edge with kubeedge

    Xiong, Y., Sun, Y., Xing, L., and Huang, Y. Extend cloud to edge with kubeedge. In 2018 IEEE/ACM Symposium on Edge Computing (SEC) (2018), IEEE, pp. 373–377

    2018

  65. [73]

    In Proceedings of the 18th ACM SIGPLAN/SIGOPS International Conference on Virtual Execution Environments (2022), pp

    Xu, Y., Wu, H., Zhang, W., and Hu, Y.Eop: efficient operator partition for deep learning inference over edge servers. In Proceedings of the 18th ACM SIGPLAN/SIGOPS International Conference on Virtual Execution Environments (2022), pp. 45–57

    2022

  66. [74]

    Towards{GPU} utilization prediction for cloud deep learning

    Yeung, G., Borowiec, D., Friday, A., Harper, R., and Garraghan, P. Towards{GPU} utilization prediction for cloud deep learning. In 12th USENIX Workshop on Hot Topics in Cloud Computing (HotCloud 20) (2020)

    2020

  67. [75]

    arXiv preprint arXiv:2306.03622 (2023)

    Yu, M., W ang, A., Chen, D., Yu, H., Luo, X., Li, Z., W ang, W., Chen, R., Nie, D., and Y ang, H.Faaswap: Slo-aware, gpu-efficient serverless inference via model swapping. arXiv preprint arXiv:2306.03622 (2023)

    2023 arXiv

  68. [76]

    F., Boutaba, R., and Hellerstein, J

    Zhang, Q., Zhani, M. F., Boutaba, R., and Hellerstein, J. L. Harmony: Dynamic heterogeneity-aware resource provisioning in the cloud. In 2013 IEEE 33rd International Conference on Distributed Computing Systems (2013), IEEE, pp. 510–519. Manuscript submitted to ACM

    2013

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.