{"id":"f31312c1-6da2-46c1-bde2-de032df49baf","arxiv_id":"2412.08294","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A GPU-cluster scheduler that co-locates DNN training jobs to reduce energy consumption, with simulation results showing up to 39% energy savings and under 3.2% runtime overhead.","lead":"This paper presents EaCO, a scheduling algorithm that shares GPUs among deep learning training jobs to cut energy use. In simulations it reports up to 39% lower cluster energy than existing schedulers, at a small runtime cost.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 39% energy reduction is computed in a simulator whose energy model and JCT prediction history both come from the same four CNN measurements; the paper never shows that production-trace jobs are representable by these profiles.","rationale":"The reader's weakest assumption correctly identifies generalization from four CNNs to production traces as the main risk. My stress-test sharpens this: the evaluation is not merely under-powered, it is circular in a specific way—the same measured profiles serve both as the simulator's energy model and as the scheduler's prediction history. This makes the 39% number an in-sample result. The paper does not describe the trace, the mapping from trace jobs to profiles, or the power model parameters, so the result cannot be reproduced or independently assessed. That said, the experimental characterization in Section 6.1 is valuable and the algorithm's underlying idea is reasonable. The appropriate verdict remains conditional acceptance with a requirement to release code/data and validate on held-out workloads. I thus agree with the reader and recommend no change to the verdict.","tokens_in":17497,"tokens_out":6180,"duration_ms":67339,"concrete_test":"Release the simulator and the production trace, then reproduce the 64-node experiment with the JCT prediction history H built from a held-out set of models (e.g., Transformer, LSTM) not among the four CNNs, while keeping the power model fixed. If the reported 39% energy reduction and <3.2% runtime increase change materially, or deadline violations appear, the headline claim depends on in-sample profiling. If the authors instead map trace jobs to the four CNN profiles, verify the mapping by running a random sample of trace jobs on the testbed and comparing predicted versus actual JCT, and report the error distribution against the deadline constraint.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 6.2 calibrates Gavel with the authors' own measurements of AlexNet, ResNet-18, ResNet-50, and VGG-16 on a single 8xV100 server; the same measurements initialize the prediction history H used in Algorithm 1 to estimate JCT for every co-located job. The paper never states which production trace is used, nor how trace jobs (which may use different architectures, batch sizes, GPU types, or parallelism strategies) are mapped to these four CNN profiles. In the 64-node experiment, EaCO's 39% energy saving comes from packing jobs even though capacity is ample, which is safe only if the JCT predictions are accurate enough to meet deadlines (Eq. 2). If trace jobs deviate from the four models, PredictJCT is unvalidated, so the claimed <3.2% runtime increase and 39% energy reduction may be artifacts of in-sample calibration. The energy saving is also dominated by server idle power via the Fan et al. CPU-utilization model [11], whose parameters are not reported. In addition, the 'first energy-aware scheduler' claim is contradicted by the paper's own reference [19], a CO2-aware DNN training scheduler.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":17834,"tokens_out":7448,"duration_ms":74576,"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":[{"comment":"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.","section":"§6.2 and Algorithm 1"},{"comment":"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.","section":"§6.1, Eq. (5)"},{"comment":"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.","section":"§6.2, Eqs. (1)-(4)"},{"comment":"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.","section":"§6.1, Figure 1 and Tables 1-4"},{"comment":"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.","section":"Algorithm 1"},{"comment":"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.","section":"Abstract, §1, and §7"}],"minor_comments":[{"comment":"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.","section":"Eq. (1)"},{"comment":"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.","section":"§6.2"},{"comment":"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.","section":"Figure 1"},{"comment":"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.","section":"Algorithm 2"},{"comment":"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.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The 'first energy-aware scheduler' claim is likely to draw scrutiny from reviewers and readers because the authors' own reference [19] is a CO2-aware DNN training scheduler. The paper may also need a clearer statement of novelty relative to that prior work. Given the in-sample evaluation and the unspecified simulation parameters, I would not recommend acceptance without a substantial revision and validation effort."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is reasonable: use GPU context switching to co-locate DNN training jobs, predict JCT with a mix of experimental profiles and live early-epoch feedback, and only pack when deadlines are met. The experimental study of four CNNs on an 8xV100 node is plausibly executed — nvidia-smi plus a CPU-utilization server power model is a standard approach, and the numbers on energy/JCT under co-location are a useful data point for the community. That part deserves credit.\n\nThe soft spots are real, though. The abstract and intro call EaCO the first energy-aware scheduler for DLT workloads, but reference [19] is the authors' own CO2-aware DNN training scheduler. That overclaim should be fixed. More importantly, the main result — up to 39% energy reduction — comes entirely from a Gavel simulator that is calibrated with the same four-model measurements that also initialize EaCO's prediction history. The production trace used is never described, how trace jobs map to AlexNet/ResNet/VGG profiles is never stated, and no error bars or repetitions are reported. The stress-test note is on target: in the 64-node case, EaCO packs jobs when capacity is ample, and that is safe only if PredictJCT is accurate for jobs outside the four profiles. As published, the 39% figure could be an artifact of in-sample calibration. The Fan et al. power model parameters are also missing, so the energy numbers are hard to reproduce.\n\nThat said, this is not a fundamentally broken idea. The algorithm itself is new in its specific combination of an energy objective, deadline constraints, historical prediction, and early-stage observation. The weaknesses are addressable: provide the simulator code, describe the trace and the job-to-profile mapping, compare against [19] and perhaps other energy-aware schedulers, and report variance. I would not reject it on principle.\n\nThis paper is for systems researchers working on GPU cluster scheduling and data center energy efficiency. It deserves a serious referee, but the referee should ask for substantial revision of the evaluation before the energy savings claim is taken at face value.","headline":"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].","tokens_in":18257,"tokens_out":2412,"would_cite":false,"duration_ms":28536,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sharing GPUs among DNN training jobs reduces cluster energy by up to 39% while keeping runtime increases under 3.2%.","keywords":["GPU sharing","DNN training","energy-aware scheduling","job co-location","GPU cluster","job completion time","resource utilization","context switching"],"falsifier":"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.","tokens_in":17255,"feed_emoji":"⚡","tokens_out":10213,"duration_ms":99259,"temperature":0.7,"pith_summary":"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.","feed_headline":"GPU sharing cuts DNN training energy up to 39%","feed_subtitle":"Co-locating deep-learning jobs on shared GPUs raises utilization to 97% while keeping runtime overhead under 3.2%","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"supplies the cluster simulator into which EaCO is integrated and on which the 28-node and 64-node traces are evaluated.","marker":"[3, 29]"},{"why":"the main packing-based scheduling baseline EaCO is compared against; it also contributed the early-feedback mechanism EaCO builds on.","marker":"[44]"},{"why":"documents the 52% average GPU utilization in production clusters, the under-utilization EaCO targets.","marker":"[22]"},{"why":"quantifies even lower median GPU utilization in another production trace, strengthening the motivation for sharing.","marker":"[42]"},{"why":"provides the CPU-utilization-based server power model used to estimate node power consumption in energy calculations.","marker":"[11]"},{"why":"supplies the per-job power aggregation formula used to compute total energy in the experiments.","marker":"[19]"},{"why":"defines occupancy and the saturation threshold that justify the utilization-based candidate filtering in EaCO.","marker":"[31]"}],"fun_headline_variants":["Co-locating DNN jobs on GPUs saves up to 39% energy","EaCO scheduler: 39% energy cut from GPU sharing","GPU co-location boosts DNN training energy efficiency by 44%","Sharing GPUs slashes DNN training energy, 3.2% runtime cost","Energy-aware GPU scheduling trims DNN training power up to 39%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Co-locating DNN jobs on GPUs saves up to 39% energy","EaCO scheduler: 39% energy cut from GPU sharing","GPU co-location boosts DNN training energy efficiency by 44%","Sharing GPUs slashes DNN training energy, 3.2% runtime cost","Energy-aware GPU scheduling trims DNN training power up to 39%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1464,"prompt_tokens":1023,"completion_tokens":441,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":339}},"tokens_in":639,"tokens_out":441,"duration_ms":4535,"temperature":1.0,"reasoning_tokens":339,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:57:59.092559+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Gandiva: Introspec- tive Cluster Scheduling for Deep Learning","cited_arxiv_id":null,"evidence_quote":"the main packing-based scheduling baseline EaCO is compared against; it also contributed the early-feedback mechanism EaCO builds on."},{"cited_title":"Analysis of Large- Scale Multi-Tenant GPU Clusters For DNN Training Workloads","cited_arxiv_id":null,"evidence_quote":"documents the 52% average GPU utilization in production clusters, the under-utilization EaCO targets."},{"cited_title":"MLaaS in the Wild: Workload Analysis and Scheduling in Large-Scale Heterogeneous GPU Clusters","cited_arxiv_id":null,"evidence_quote":"quantifies even lower median GPU utilization in another production trace, strengthening the motivation for sharing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the CPU-utilization-based server power model used to estimate node power consumption in energy calculations."},{"cited_title":"CO2 Emission Aware Scheduling for Deep Neural Network Training Workloads","cited_arxiv_id":null,"evidence_quote":"supplies the per-job power aggregation formula used to compute total energy in the experiments."},{"cited_title":"CUDA C Best Practices Guide - Oc- cupancy","cited_arxiv_id":null,"evidence_quote":"defines occupancy and the saturation threshold that justify the utilization-based candidate filtering in EaCO."}],"review_version":1}