Pith. sign in

REVIEW 6 major objections 5 minor 49 references

EaCO: Resource Sharing Dynamics and Its Impact on Energy Efficiency for DNN Training

T0 review · 6 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Sharing GPUs among DNN training jobs reduces cluster energy by up to 39% while keeping runtime increases under 3.2%.

desk verdict Sensible energy-aware co-allocation idea, but the headline 39% energy saving is an in-sample simulator result and the 'first' claim collides with the authors' own reference [19]. read the letter →

arxiv 2412.08294 v1 pith:HQAZUG7C submitted 2024-12-11 cs.DC

Kawsar Haghshenas , Mona Hashemi This is my paper · ORCID
classification cs.DC
keywords GPUsharingDNNtrainingenergy-awareschedulingjobco-locationclustercompletiontimeresourceutilizationcontextswitching
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 argues that sharing GPUs among multiple deep-learning training jobs can substantially reduce data-center energy consumption without meaningful performance loss, and it presents a scheduling algorithm, EaCO, designed to realize that gain. In experiments on a server with eight GPUs, co-locating pairs and triples of CNN training jobs cuts energy by 30–44% for individual job sets and raises average GPU utilization to as high as 97%, at the price of a 3–19% increase in job completion time. In cluster-scale simulations using production traces, EaCO reduces total energy by up to 39% compared with existing schedulers while increasing average job runtime by less than 3.2%. The paper's central claim is that hardware-supported context switching makes co-location cheap enough that energy efficiency can become a first-class scheduling objective, as long as placements are checked against job deadlines.

What carries the argument

The load-bearing mechanism is hardware-supported GPU context switching, which lets multiple training jobs' kernels time-share a GPU with small switching overhead, turning otherwise idle SMs into useful work. On top of that, EaCO's machinery is a hybrid prediction-and-correction loop: FindCandidates selects GPU sets whose core and memory utilization are below thresholds; PredictJCT estimates job completion times from experimental and historical data; and an early-stage observation phase profiles the co-located jobs for one epoch, finalizing the placement only if all deadlines remain feasible and otherwise undoing it at an epoch boundary. The algorithm minimizes an alpha-weighted objective combining total energy and average time per epoch, with job deadlines as hard constraints.

What would settle it

Run a multi-node trace containing transformer, recommendation, and graph-neural-network training jobs while measuring node power directly with hardware power meters; if EaCO's total-energy reduction falls well short of 39% relative to the same baselines, or its runtime increase exceeds the claimed 3.2%, the calibration assumption does not generalize.

Watch

Extended reading notes

Core claim

The central discovery is that GPU co-location is energy-positive for DNN training: when several training jobs share a GPU, the combined energy is lower than the sum of their exclusive executions because idle compute capacity is filled while the node's base power is paid once. EaCO turns this into a scheduler by predicting job completion time for each candidate co-location from a history of experimental measurements, placing a job on the highest-utilization eligible GPUs, and then observing one full epoch (one pass through the training data) of all co-located jobs before finalizing the placement; if a deadline is at risk, the allocation is reversed at an epoch boundary. The scheduling objective is a weighted sum of total energy and average time per epoch, subject to per-job deadlines and utilization and memory thresholds. The paper states these mechanisms produce up to 44% energy savings per job set in single-node experiments and up to 39% total energy savings in 28-node and 64-node cluster simulations.

Load-bearing premise

The paper assumes that the energy and interference behavior measured for four CNN training jobs on one server generalizes to the diverse jobs in the production traces and to the 64-node cluster, so the same 39% savings appear at scale.

Editorial extensions

If this is right

  • If the 39% cluster-level energy reduction holds, GPU clusters can cut a large share of their training-related energy without waiting for more efficient hardware.
  • The under-3.2% runtime increase means energy-aware co-location can be adopted without violating typical batch-training deadlines.
  • The 97% GPU utilization observed in co-location experiments implies that existing clusters have substantial headroom to absorb more jobs before buying new GPUs.
  • Because EaCO reduces the number of active nodes by up to 47%, operators can combine it with low-power idle states to compound the energy savings.

Reading between the lines

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

  • Editorial extension: the epoch-to-epoch utilization consistency the paper measures could also drive runtime adaptation, such as adjusting batch size or GPU clock speed, not just placement decisions, because the same early signals predict steady-state behavior.
  • Editorial extension: if the underlying co-location energy model holds, combining EaCO's placement decisions with carbon-aware timing could shift energy-intensive training to periods of low grid carbon intensity, multiplying the environmental benefit.
  • Editorial extension: the evaluation's four CNN models leave open how memory-heavy transformer or recommendation workloads behave under co-location; a direct test of EaCO on those workloads would sharpen the generalization claim.
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

6 major / 5 minor

Summary. The paper studies energy-efficient co-scheduling of DNN training jobs on GPU clusters. It contributes experimental measurements of co-location effects for four CNN models on one 8xV100 server, and it proposes EaCO, a greedy scheduling algorithm that selects candidate GPU sets using utilization thresholds, predicts JCT from a history initialized with the same measurements, and verifies allocations after one epoch. The evaluation combines these experiments with Gavel-based simulations on 28- and 64-node clusters and claims up to 44% per-job energy-efficiency gains and up to 39% cluster energy reduction with less than 3.2% average JCT increase.

Significance. The motivating question is timely: GPU under-utilization is documented, and energy-aware scheduling for DLT clusters is an active concern. The paper has concrete strengths: direct GPU power measurements via nvidia-smi, a real experimental study of co-location, and the attempt to ground cluster simulation in measured data. If the claims held, EaCO would be a useful addition to cluster schedulers. However, as written the evaluation is in-sample, lacks statistical and power-model detail, and leaves key simulation parameters unspecified, so the headline numbers should be treated as preliminary rather than established results.

major comments (6)
  1. [§6.2 and Algorithm 1] The simulation evaluation is in-sample by construction. The Gavel simulator is calibrated with the authors' own measurements of AlexNet, ResNet-18, ResNet-50, and VGG-16 on a single 8xV100 server, and the history H in Algorithm 1 (Line 1) is initialized with the same experimental measurements. No production trace is identified in §6.2, and no mapping is given from trace jobs to these four architectures, batch sizes, or parallelism strategies. Consequently, the 39% energy saving and <3.2% JCT increase are claims about the four-model workload, not about the production traces mentioned in the abstract. The authors should state the trace, define the job-to-profile mapping, and add an out-of-sample validation (for example, leave-one-model-out or a held-out workload composed of other models) so that PredictJCT is evaluated on data not used to initialize it.
  2. [§6.1, Eq. (5)] The server power term P(server,t) is 'estimated based on CPU utilization' following Fan et al. [11], but the idle power, peak power, and CPU-utilization coefficient are not reported. Because server idle power dominates cluster energy when nodes are active, the 30-44% efficiency gain in §6.1 and the 39% reduction in §6.2 depend on a model whose constants are not stated. Report the parameters, and include a sensitivity analysis or a direct comparison against node-level power measurements.
  3. [§6.2, Eqs. (1)-(4)] The simulation setup is under-specified in a way that makes the results non-reproducible. No values are reported for the trade-off weight α in Eq. (1), for the thresholds U_threshold and mem_threshold in Eqs. (3)-(4), or for how job deadlines in Eq. (2) are generated. These parameters directly control EaCO's packing aggressiveness and its SLO enforcement, so different choices could substantially change the claimed 39% energy reduction and <3.2% JCT increase. A table of simulation parameters, including the deadline distribution, should be added.
  4. [§6.1, Figure 1 and Tables 1-4] All energy and JCT values are reported as single runs, without repetitions or variance. DNN training is stochastic and power draw fluctuates; the claimed 30-44% energy improvement and 3-19% JCT overhead are central quantitative results, so they should be accompanied by at least three repetitions with mean±std or confidence intervals, and an indication of hardware and software configuration stability.
  5. [Algorithm 1] The algorithm does not handle the case where no candidate GPU set in L meets the deadline condition. In the inner loop (Lines 4-11), if every G is removed from L without a successful allocation, the loop continues with an exhausted list, so accessing the 'highest utilization' set at Line 5 is undefined. In addition, if an allocation is undone at Line 18, the jump 'Go to Line 3' does not recompute L (Line 2), so the same exhausted candidate list is reused. The algorithm also has no fallback to exclusive allocation or to waiting, which is necessary for a scheduler that must always place a job. This needs a concrete fix and a trace-based check that EaCO never fails to schedule a job.
  6. [Abstract, §1, and §7] The paper repeatedly states that EaCO is 'the first energy-aware scheduling algorithm' for DLT workloads, but reference [19] (Haghshenas et al., 2022) is a CO2-aware DNN training scheduler, and §7 does not position EaCO relative to it. CO2-aware scheduling is energy-aware in its objective, so this novelty claim is too strong and should be revised or substantiated with a clear distinction.
minor comments (5)
  1. [Eq. (1)] The objective function is typeset ambiguously: 'Minimize α. ∑ E_j + (1 − α).AvgT PE' should be written with explicit parentheses and operators, for example as α * Σ_j E_j + (1 − α) * AvgTPE.
  2. [§6.2] The workload trace used for simulation is never named; if it is a public trace (for example, Philly or Alibaba), cite it and describe any preprocessing and filtering.
  3. [Figure 1] The numeric labels over the bars (for example, '1.4 1.5 1.5 1.8 1.7') are cryptic and are not explained in the caption; either remove them or add a legend that identifies what they represent.
  4. [Algorithm 2] Line 1 of Algorithm 2 has an unmatched parenthesis and an imprecise description: 'All subsets of GPUs based on ( j's requested GPUs' should be reworded to clarify whether subsets may span multiple nodes and how the requested GPU count determines the set size.
  5. [Table 3] The four-job row reports no average epoch time, but §6.1 later discusses the JCT behavior for that configuration; adding the measurement or explaining why it is unavailable would improve completeness.

Circularity Check

1 steps flagged · score 6.0 of 10

In-sample simulation: EaCO's JCT 'predictions' and the simulator's energy/perf models are populated from the same four CNN measurements, so the 39% energy saving replays the measured co-location saving rather than independently predicting it.

  1. fitted input called prediction [Algorithm 1, Lines 1, 6, 13-14; Section 6.2, paragraph 1]
    "Initialize history H with experimental measurements. / Predict JCT for all co-located jobs on G and j using H. / we (1) integrated the empirical data and measurements derived from our experiments on power consumption and performance into the simulator."

    Both H (Algorithm 1) and the simulator's added energy/perf modules (Section 6.2) are populated from the same four-model experiments (Tables 1-4). For any profiled co-location, 'Predict JCT' returns the measured JCT from Table 3 and energy uses the measured powers, so the prediction is an input lookup by construction. The 39% energy reduction (Figure 3) is thus the measured 30-44% co-location saving replayed through a packing schedule, not an independent validation. No trace-to-model mapping is given, so the simulation does not validate predictions for unseen workloads.

full rationale

The central quantitative claim is a simulation result whose inputs and 'predictions' draw on the identical set of author-run measurements. Algorithm 1 initializes history H with experimental measurements, and Section 6.2 explicitly integrates the same empirical data into the Gavel simulator, so PredictJCT is a table lookup for the profiled model combinations. Consequently, the 39% energy saving is an in-sample replay of the measured 30-44% co-location savings; the deadline constraint (Eq. 2) only filters placements, while the cost and performance values are input lookups. This is a partial circularity (score 6): the measured experimental facts and the external CPU-utilization power model are independent inputs, and the scheduler design is not forced by a self-citation chain, but the headline 'prediction' reduces to the calibration data by construction. The paper's separate novelty claim of being the 'first energy-aware' scheduler is contradicted by its own reference [19], though that is a factual/correctness issue rather than a circularity and is not scored here.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest heavily on experimentally measured values that are reused in the simulator and in the algorithm's prediction history. No new physical entities are introduced. The free parameters are scheduling thresholds and the alpha trade-off, which are chosen by hand and not justified. The key domain assumptions are the representativeness of early epochs, the accuracy of the CPU-based power model, and the generalization of four model measurements to production workloads.

free parameters (3)
  • alpha (trade-off weight in cost function) = not specified
    Equation (1) sets a weighted sum of total energy and average time per epoch. The paper does not describe how alpha is chosen or tuned, but it directly controls the balance between energy and performance.
  • U_threshold (GPU core utilization threshold) = not specified
    Algorithm 2 filters candidate GPUs with core utilization below U_threshold. The paper never gives the numerical value used in experiments or simulations.
  • mem_threshold (memory utilization threshold) = not specified
    Algorithm 2 also filters by memory utilization against mem_threshold. The value is not reported, and it affects which GPUs are eligible for sharing.
assumptions (4)
  • domain assumption Early-stage (first-epoch) resource utilization and epoch time are representative of the entire training job's steady-state behavior.
    Algorithm 1 lines 12-18 finalize an allocation after observing one epoch per co-located job, assuming this observation predicts future JCT and deadlines reliably.
  • domain assumption The server power model based on CPU utilization (from Fan et al. [11]) plus GPU power from nvidia-smi accurately captures total node energy.
    Equation (5) defines job power as server power plus GPU power. This imported model is used in all experiments and in the simulator, but is not validated against whole-node power meters.
  • domain assumption GPU hardware context switching allows multiple DNN training jobs to share the same GPU without failures and with interference levels matching the four CNN models tested.
    The entire EaCO approach relies on safe co-location; the experiments cover only one server type and four image classification models, yet the simulation assumes these measurements generalize to arbitrary DLT jobs.
  • domain assumption The Gavel simulator, after integrating the authors' measured data, faithfully represents large-scale cluster behavior for production traces.
    Section 6.2 states the simulator was modified for energy evaluation, but no validation against a real cluster is provided, so the 28-node and 64-node simulation results depend on this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EaCO: Resource Sharing Dynamics and Its Impact on Energy Efficiency for DNN Training." pith.science (2026). https://pith.science/paper/HQAZUG7C

@misc{pith2026241208294,
  author       = {Pith},
  title        = {Pith review of: EaCO: Resource Sharing Dynamics and Its Impact on Energy Efficiency for DNN Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HQAZUG7C}},
  note         = {Machine review of arXiv:2412.08294}
}
read the original abstract

Deep Learning Training (DLT) is a growing workload in shared GPU/CPU clusters due to its high computational cost and increasing number of jobs. This contributes to significant energy consumption in GPU clusters, further exacerbated by GPU under-utilization, as shown in production cluster logs. Addressing this challenge requires workload scheduling and resource allocation policies for efficient GPU sharing to improve resource and energy efficiency while maintaining performance. However, previous works primarily optimize for performance, often overlooking or even sacrificing energy efficiency. In this paper, we present EaCO, the first energy-aware scheduling algorithm designed specifically for DLT workloads in GPU clusters. EaCO leverages hardware-supported context switching to enable GPU sharing across multiple DLT jobs, improving resource and energy utilization. GPU sharing can increase Job Completion Time (JCT) and may lead to contention if not employed carefully. To address this, EaCO integrates experiment and historical-based predictions as well as early-stage observations, ensuring performance expectations are met while optimizing energy efficiency. We begin by experimentally exploring the dynamics of co-locating DLTs, investigating its impact on energy and resource utilization. Our results show that co-location improves energy efficiency by up to 44% for individual jobs, and increases average GPU utilization to as high as 97%. Additionally, evaluations on large-scale clusters using production traces demonstrate that EaCO reduces total energy by up to 39% compared to existing algorithms, which comes with a minimal increase in job runtime-less than 3.2% in our simulations.

Figures

Figures reproduced from arXiv: 2412.08294 by the authors.

Figure 1
Figure 1. Total energy and average JCT for running a set of [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Resource utilization (GPU, CPU, and memory), while running different job combinations. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Total energy and average job runtime for executing [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Number of active nodes employing different algo [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 45 canonical work pages

  1. [19]

    CO2 Emission Aware Scheduling for Deep Neural Network Training Workloads

    HAGHSHENAS , K., S ETZ , B., AND AIELLO , M. CO2 Emission Aware Scheduling for Deep Neural Network Training Workloads. In 2022 IEEE International Con- ference on Big Data (Big Data) (2022), IEEE, pp. 1542– 1549

    work page 2022

  2. [11]

    FAN, X., WEBER , W.-D., AND BARROSO , L. A. Power Provisioning for A Warehouse-Sized Computer. ACM SIGARCH computer architecture news 35, 2 (2007), 13– 23

    work page 2007

  3. [1]

    https://developer

    NVIDIA Nsight Compute. https://developer. nvidia.com/nsight-compute

    work page

  4. [2]

    https://developer.nvidia.com/ nvidia-system-management-interface

    NVIDIA System Management Inter- face. https://developer.nvidia.com/ nvidia-system-management-interface

    work page

  5. [3]

    https://github.com/ stanford-futuredata/gavel, 2020

    Gavel. https://github.com/ stanford-futuredata/gavel, 2020

    work page 2020

  6. [4]

    Topology-Aware GPU Schedul- ing for Learning Workloads in Cloud Environments

    AMARAL , M., P OLO , J., C ARRERA , D., S EELAM , S., AND STEINDER , M. Topology-Aware GPU Schedul- ing for Learning Workloads in Cloud Environments. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Anal- ysis (2017), pp. 1–12

    work page 2017

  7. [5]

    DNNAbacus: Toward Accurate Computational Cost Prediction for Deep Neural Networks

    BAI, L., J I, W., L I, Q., Y AO, X., X IN, W., AND ZHU, W. DNNAbacus: Toward Accurate Computational Cost Prediction for Deep Neural Networks. arXiv preprint arXiv:2205.12095 (2022)

    arXiv 2022

  8. [6]

    PipeSwitch: Fast Pipelined Context Switching for Deep Learning Applications

    BAI, Z., Z HANG , Z., Z HU, Y., AND JIN, X. PipeSwitch: Fast Pipelined Context Switching for Deep Learning Applications. In 14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20) (2020), pp. 499–514

    work page 2020

Show all 49 references
  1. [7]

    Online Job Scheduling in Distributed Machine Learning Clusters

    BAO, Y., P ENG , Y., W U, C., AND LI, Z. Online Job Scheduling in Distributed Machine Learning Clusters. In IEEE INFOCOM 2018-IEEE Conference on Com- puter Communications (2018), IEEE, pp. 495–503

    2018

  2. [8]

    Borg, omega, and kubernetes

    BURNS , B., G RANT , B., O PPENHEIMER , D., B REWER , E., AND WILKES , J. Borg, omega, and kubernetes. Communications of the ACM 59, 5 (2016), 50–57. 12

    2016

  3. [9]

    Balancing Efficiency and Fairness in Heterogeneous GPU Clusters for Deep Learning

    CHAUDHARY , S., R AMJEE , R., S IVATHANU , M., K WA- TRA , N., AND VISWANATHA , S. Balancing Efficiency and Fairness in Heterogeneous GPU Clusters for Deep Learning. In Proceedings of the Fifteenth European Conference on Computer Systems (2020), pp. 1–16

    2020

  4. [10]

    Imagenet: A Large-Scale Hierarchical Image Database

    DENG , J., D ONG , W., S OCHER , R., L I, L.-J., L I, K., AND FEI-F EI, L. Imagenet: A Large-Scale Hierarchical Image Database. In 2009 IEEE conference on computer vision and pattern recognition (2009), Ieee, pp. 248– 255

    2009

  5. [12]

    Runtime Performance Prediction for Deep Learn- ing Models with Graph Neural Network

    GAO, Y., G U, X., Z HANG , H., L IN, H., AND YANG , M. Runtime Performance Prediction for Deep Learn- ing Models with Graph Neural Network. In 2023 IEEE/ACM 45th International Conference on Software Engineering: Software Engineering in Practice (ICSE- SEIP) (2023), IEEE, pp. 368–380

    2023

  6. [13]

    Estimating GPU Memory Con- sumption of Deep Learning Models

    GAO, Y., L IU, Y., Z HANG , H., L I, Z., Z HU, Y., L IN, H., AND YANG , M. Estimating GPU Memory Con- sumption of Deep Learning Models. In Proceedings of the 28th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Founda- tions of Software Engin...

    2020

  7. [14]

    Google Cloud GPU

    GOOGLE CLOUD . Google Cloud GPU. https:// cloud.google.com/gpu

  8. [15]

    Google Cloud AI Platform Training and Prediction - SLA

    GOOGLE CLOUD . Google Cloud AI Platform Training and Prediction - SLA. https://cloud.google.com/ ai-platform/training-and-prediction/sla , 2020

    2020

  9. [16]

    G., Z HU, Y., JEON , M., Q IAN , J., L IU, H., AND GUO, C

    GU, J., C HOWDHURY , M., S HIN , K. G., Z HU, Y., JEON , M., Q IAN , J., L IU, H., AND GUO, C. Tiresias: A GPU Cluster Manager for Distributed Deep Learning. In 16th USENIX Symposium on Networked Systems Design and Implementation (NSDI 19) (2019), pp. 485–500

    2019

  10. [17]

    Liquid: In- telligent Resource Estimation and Network-Efficient Scheduling for Deep Learning Jobs on Distributed GPU Clusters

    GU, R., C HEN , Y., L IU, S., D AI, H., C HEN , G., ZHANG , K., C HE, Y., AND HUANG , Y. Liquid: In- telligent Resource Estimation and Network-Efficient Scheduling for Deep Learning Jobs on Distributed GPU Clusters. IEEE Transactions on Parallel and Distributed Systems 33, 11 ...

    2021

  11. [18]

    Magnetic: Multi- agent machine learning-based approach for energy ef- ficient dynamic consolidation in data centers

    HAGHSHENAS , K., P AHLEVAN , A., Z APATER , M., MOHAMMADI , S., AND ATIENZA , D. Magnetic: Multi- agent machine learning-based approach for energy ef- ficient dynamic consolidation in data centers. IEEE Transactions on Services Computing 15, 1 (2019), 30– 44

    2019

  12. [20]

    Deep Resid- ual Learning for Image Recognition

    HE, K., ZHANG , X., REN, S., AND SUN, J. Deep Resid- ual Learning for Image Recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (2016), pp. 770–778

    2016

  13. [21]

    C., LIN, X., AND DENG , N

    JAJOO , A., H U, Y. C., LIN, X., AND DENG , N. SLearn: A Case for Task Sampling Based Learning for Cluster Job Scheduling. IEEE Transactions on Cloud Comput- ing 11, 3 (2022), 2664–2680

    2022

  14. [22]

    Analysis of Large- Scale Multi-Tenant GPU Clusters For DNN Training Workloads

    J EON , M., V ENKATARAMAN , S., P HANISHAYEE , A., QIAN , J., X IAO, W., AND YANG , F. Analysis of Large- Scale Multi-Tenant GPU Clusters For DNN Training Workloads. In 2019 USENIX Annual Technical Confer- ence (USENIX ATC 19) (2019), pp. 947–960

    2019

  15. [23]

    KRIZHEVSKY , A., S UTSKEVER , I., AND HINTON , G. E. ImageNet Classification with Deep Convolutional Neural Networks. Advances in neural information pro- cessing systems 25 (2012), 1097–1105

    2012

  16. [24]

    N., S UN, X., C HOWDHURY , M., AND LIU, Z

    LE, T. N., S UN, X., C HOWDHURY , M., AND LIU, Z. AlloX: Compute Allocation in Hybrid Clusters. In Pro- ceedings of the Fifteenth European Conference on Com- puter Systems (2020), pp. 1–16

    2020

  17. [25]

    Deep Learning

    LECUN, Y., B ENGIO , Y., AND HINTON , G. Deep Learning. nature 521, 7553 (2015), 436–444

    2015

  18. [26]

    LIU, H., L IU, S., W EN, C., AND WONG , W. E. TBEM: Testing-Based GPU-Memory Consumption Estimation for Deep Learning. IEEE Access 10 (2022), 39674– 39680

    2022

  19. [27]

    Themis: Fair and Efficient GPU Cluster Scheduling

    MAHAJAN , K., B ALASUBRAMANIAN , A., S INGHVI , A., V ENKATARAMAN , S., A KELLA , A., P HAN - ISHAYEE , 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

  20. [28]

    Azure Databricks Documentation - GPU Clusters

    MICROSOFT . Azure Databricks Documentation - GPU Clusters. https://learn.microsoft.com/nl-nl/ azure/databricks/clusters/gpu, 2024

    2024

  21. [29]

    Heterogeneity-Aware Cluster Scheduling Policies for Deep Learning Workloads

    NARAYANAN , D., S ANTHANAM , K., K AZHAMI - AKA , F., P HANISHAYEE , 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. 13

    2020

  22. [30]

    CUDA C Best Practices Guide - Calculating Occupancy

    NVIDIA . CUDA C Best Practices Guide - Calculating Occupancy. https://docs.nvidia.com/cuda/ cuda-c-best-practices-guide/index.html# calculating-occupancy, 2023

    2023

  23. [31]

    CUDA C Best Practices Guide - Oc- cupancy

    NVIDIA . CUDA C Best Practices Guide - Oc- cupancy. https://docs.nvidia.com/cuda/ cuda-c-best-practices-guide/index.html# occupancy, 2023

    2023

  24. [32]

    NVIDIA Data Center GPU Manager (DCGM)

    NVIDIA C ORPORATION . NVIDIA Data Center GPU Manager (DCGM). https://developer.nvidia. com/dcgm

  25. [33]

    AI and Compute

    OPENAI. AI and Compute. https://openai.com/ research/ai-and-compute, 2018

    2018

  26. [34]

    Carbon emissions and large neu- ral network training

    PATTERSON , D., G ONZALEZ , J., L E, Q., L IANG , C., MUNGUIA , L.-M., R OTHCHILD , D., S O, D., T EXIER , M., AND DEAN , J. Carbon emissions and large neu- ral network training. arXiv preprint arXiv:2104.10350 (2021)

    2021 arXiv

  27. [35]

    PyTorch ImageNet Ex- amples

    PYTORCH CONTRIBUTORS . PyTorch ImageNet Ex- amples. https://github.com/pytorch/examples/ tree/main/imagenet, 2023

    2023

  28. [36]

    Very Deep Con- volutional Networks for Large-Scale Image Recognition

    SIMONYAN , K., AND ZISSERMAN , A. Very Deep Con- volutional Networks for Large-Scale Image Recognition. arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  29. [37]

    Multifactor Priority Plugin

    SLURM . Multifactor Priority Plugin. https://slurm. schedmd.com/priority_multifactor.html, 2023

    2023

  30. [38]

    En- ergy Aware Consolidation for Cloud Computing

    SRIKANTAIAH , S., K ANSAL , A., AND ZHAO, F. En- ergy Aware Consolidation for Cloud Computing. In USENIX HotPower’08: Workshop on Power Aware Com- puting and Systems at OSDI (2008)

    2008

  31. [39]

    Orion: Interference-Aware, Fine-grained GPU Sharing for ML Applications

    STRATI , F., M A, X., AND KLIMOVIC , A. Orion: Interference-Aware, Fine-grained GPU Sharing for ML Applications. In Proceedings of the Nineteenth Euro- pean Conference on Computer Systems(2024), pp. 1075– 1092

    2024

  32. [40]

    Energy and Policy Considerations for Modern Deep Learning Research

    STRUBELL , E., G ANESH , A., AND MCCALLUM , A. Energy and Policy Considerations for Modern Deep Learning Research. In Proceedings of the AAAI confer- ence on artificial intelligence(2020), vol. 34, pp. 13693– 13696

    2020

  33. [41]

    K., M URTHY, A

    VAVILAPALLI , V. K., M URTHY, A. C., D OUGLAS , C., AGARWAL , S., K ONAR , M., E VANS, R., G RAVES , T., LOWE, J., SHAH , H., SETH , S., ET AL . Apache Hadoop YARN: Yet Another Resource Negotiator. InProceed- ings of the 4th annual Symposium on Cloud Computing (2013), pp. 1–16

    2013

  34. [42]

    MLaaS in the Wild: Workload Analysis and Scheduling in Large-Scale Heterogeneous GPU Clusters

    WENG , Q., X IAO, W., Y U, Y., WANG , W., WANG , C., HE, J., L I, Y., Z HANG , L., L IN, 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 (NS...

    2022

  35. [43]

    Transparent GPU Sharing in Container Clouds for Deep Learning Workloads

    WU, B., Z HANG , Z., B AI, Z., L IU, X., AND JIN, X. Transparent GPU Sharing in Container Clouds for Deep Learning Workloads. In 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI

  36. [44]

    Gandiva: Introspec- tive Cluster Scheduling for Deep Learning

    XIAO, W., B HARDWAJ , R., R AMJEE , R., S IVATHANU , M., K WATRA, N., H AN, Z., P ATEL, P., P ENG , X., ZHAO, H., Z HANG , Q., ET AL . Gandiva: Introspec- tive Cluster Scheduling for Deep Learning. In 13th USENIX Symposium on Operating Systems Design and Implementation (OSDI 1...

    2018

  37. [45]

    AntMan: Dynamic Scaling on GPU Clusters for Deep Learning

    XIAO, W., R EN, S., L I, Y., Z HANG , Y., H OU, P., L I, Z., F ENG , Y., LIN, W., AND JIA, Y. AntMan: Dynamic Scaling on GPU Clusters for Deep Learning. In 14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20) (2020), pp. 533–548

    2020

  38. [46]

    B., J ETTE , M

    YOO, A. B., J ETTE , M. A., AND GRONDONA , M. Slurm: Simple linux utility for resource management. In Workshop on job scheduling strategies for parallel processing (2003), Springer, pp. 44–60

    2003

  39. [47]

    Salus: Fine-Grained GPU Sharing Primitives for Deep Learning Applica- tions

    YU, P., AND CHOWDHURY , M. Salus: Fine-Grained GPU Sharing Primitives for Deep Learning Applica- tions. arXiv preprint arXiv:1902.04610 (2019)

    2019 arXiv

  40. [48]

    ZHANG , H., S TAFMAN , L., O R, A., AND FREEDMAN , M. J. SLAQ: Quality-Driven Scheduling for Distributed Machine Learning. In Proceedings of the 2017 Sympo- sium on Cloud Computing (2017), pp. 390–404

    2017

  41. [49]

    An Empirical Study on Program Fail- ures of Deep Learning Jobs

    ZHANG , R., X IAO, W., Z HANG , H., L IU, Y., L IN, H., AND YANG , M. An Empirical Study on Program Fail- ures of Deep Learning Jobs. In Proceedings of the ACM/IEEE 42nd International Conference on Software Engineering (2020), pp. 1159–1170. 14

    2020

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.