{"id":"7bb52201-1032-449c-a97f-3cbee46e3ff3","arxiv_id":"2506.04902","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A TOPSIS-based Kubernetes scheduler reduces scheduling-decision energy by up to 39.1% in a small heterogeneous GKE cluster, though the broader cluster-wide energy impact is extrapolated from assumed parameters.","lead":"GreenPod, a TOPSIS-based scheduler for Kubernetes, claims to cut scheduling-energy consumption by up to 39.1% compared to the default scheduler in a heterogeneous Google cluster. The paper also estimates annual carbon and cost savings for data centers, but those estimates rest on a questionable extrapolation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section V.E applies scheduling-only energy savings (19.38%) to total job energy (0.024 kWh), conflating scheduler overhead with workload energy; the MWh, CO2, and cost extrapolations are unsupported.","rationale":"The reader's weakest_assumption correctly identifies the load-bearing flaw: Section V.E transfers a scheduling-decision energy optimization percentage to total job energy without justification. My independent check confirms the quantitative severity: the measured scheduling energy is roughly 0.5 kJ per decision, while the extrapolation's per-job total energy is 0.024 kWh (86.4 kJ), so a 19.38% saving on the former is only about 0.1% of the latter. This invalidates the headline real-world impact claims (MWh, CO2, cost, carbon credits) while leaving only the narrow and much weaker claim that GreenPod reduces scheduling-decision energy. The paper does present a plausible TOPSIS scheduler and reproducible artifact reference, and the narrow scheduling-energy measurement may be sound, but the central advertised contributions are the energy, carbon, and cost savings, and those are not supported by the reported data. The abstract's inconsistent carbon figure (3.39 vs 3.99 metric tons) and lack of error bars further reduce confidence. Given that the main impact claims are overstated by roughly two orders of magnitude, rejection of the current version is appropriate; a revision that reframes the claims to scheduling overhead or measures total workload energy could warrant reconsideration.","tokens_in":10706,"tokens_out":2244,"duration_ms":28373,"concrete_test":"Recompute the impact analysis using only measured scheduling energy: multiply the average per-decision savings from Table VI (0.0886 kJ) by the number of scheduling decisions required for 6,304 daily jobs, and compare the resulting daily savings with Section V.E's 0.0293 MWh. Alternatively, run the GreenPod and default schedulers on the same GKE cluster while measuring total cluster energy over complete workload executions; if the total-energy saving is approximately 0.1% rather than 19.38%, the Section V.E extrapolation and any application of the 39.1% headline to workload energy are invalid.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's own definitions and Table VI indicate that the measured 'Energy Consumed (kJ)' quantifies the energy of scheduling decisions, not the total energy of the executed workload. The 19.38% average optimization is therefore a reduction in scheduling-decision energy. Section V.E then multiplies a per-job total-energy estimate from the Dayarathna blade-server power model (0.024 kWh/job) by 0.1938 to claim 0.0293 MWh/day and 10.70 MWh/year per cluster. This step implicitly assumes that the 19.38% scheduling-energy improvement transfers to total job energy, an assumption nowhere stated or justified. The paper's own Table VI shows default scheduling energy around 0.5036 kJ (0.00014 kWh) per scheduling decision; the average savings of 0.0886 kJ is about 0.000025 kWh per decision, roughly 0.1% of the 0.024 kWh per-job estimate—not 19.38%. Thus the extrapolated annual savings, CO2 reductions, and cost benefits are inflated by about two orders of magnitude. The abstract's 3.39 metric tons versus the body's 3.99 metric tons is a further inconsistency, but the core problem is the unit mismatch in Section V.E.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents GreenPod, a TOPSIS-based Kubernetes scheduler that ranks candidate nodes on five weighted criteria: execution time, energy consumption, core availability, memory availability, and resource balance. The authors evaluate GreenPod against the default Kubernetes scheduler on a heterogeneous Google Kubernetes Engine cluster under three competition levels and four weighting schemes, reporting energy savings of up to 39.1%. They then extrapolate these savings to a production-scale cluster using SURF Lisa job statistics and a blade-server power model, claiming annual savings of 10.70 MWh, 3.99 metric tons of CO2, and about $1,380 per cluster. The paper also releases an open-source scheduler. The core experimental comparison is real, but the headline claim and the real-world impact analysis rest on a conflation between scheduling-decision energy and total workload energy, which is a load-bearing error.","tokens_in":10889,"tokens_out":3694,"duration_ms":44402,"significance":"If the experimental results were properly scoped to scheduling-decision energy only, GreenPod would be a modest engineering contribution: it demonstrates that a TOPSIS-based custom scheduler can be integrated into Kubernetes and can reduce the energy spent in scheduling decisions in a heterogeneous cluster. However, the paper's stated significance—energy-efficient orchestration for AIoT workloads with meaningful cluster-wide savings—depends entirely on the unsupported extrapolation in Section V.E. The paper explicitly defines its measured 'Energy Consumed (kJ)' as quantifying the efficiency of scheduling decisions (Table IV), yet the abstract and conclusion make unqualified claims about reducing energy consumption, and the impact analysis applies the scheduling-decision savings percentage to total per-job energy. Because the extrapolation inflates the savings by roughly two orders of magnitude, the environmental and economic claims (MWh, CO2, cost, carbon credits) are not credible as presented. The underlying experimental comparison may be salvageable, but only with a major reframing of the claims and new measurements of total workload energy.","major_comments":[{"comment":"Section V.E contains a unit mismatch that undermines the real-world impact claims. Table IV defines 'Energy Consumed (kJ)' as quantifying the efficiency of scheduling decisions, and Table VI reports values around 0.5 kJ per decision for the default scheduler. The average optimization of 19.38% is therefore the reduction in scheduling-decision energy, which is roughly 0.0886 kJ (about 0.000025 kWh) per decision. Section V.E multiplies a per-job total-energy estimate of 0.024 kWh (derived from the Dayarathna blade-server power model with assumed parameters) by 0.1938 to obtain daily savings of 0.0293 MWh and annual savings of 10.70 MWh. This assumes, without any justification, that the 19.38% scheduling-decision energy improvement transfers to the total energy consumed by the job. The paper's own numbers show that scheduling-decision energy is about 0.1% of the 0.024 kWh per-job estimate, so the extrapolated MWh, CO2, and cost savings are inflated by approximately two orders of magnitude. This error is load-bearing because the paper's conclusion and contribution claims are based on these extrapolated savings.","section":"Table VI"},{"comment":"The abstract and conclusion state that GreenPod 'reduces energy consumption by up to 39.1% compared to the default scheduler.' Since the experimental metric in Table VI is explicitly defined in Table IV as 'energy consumed' by scheduling decisions, the claim should be scoped to 'scheduling-decision energy' or similar. As written, the unqualified phrasing overstates the result by implying a 39.1% reduction in total workload or cluster energy. This is not merely a wording issue: the unsupported extrapolation in Section V.E is a direct consequence of treating the scheduling-only percentage as a workload-level energy saving.","section":"V.E"},{"comment":"The paper does not specify how the energy consumption values in Table VI were measured. Section III.A mentions 'monitoring agents that collect fine-grained energy data via hardware interfaces or calibrated power models,' but Section IV never states which method was used, what hardware or sensors were involved, how the energy of a single scheduling decision was isolated from the energy of the running workloads, or whether any statistical replication was performed. Without this information, the numerical values in Table VI (e.g., identical 0.5036 kJ for all low-competition default-scheduler rows) cannot be independently assessed, which is a serious problem for the paper's central empirical claim.","section":"IV.C"}],"minor_comments":[{"comment":"The abstract reports CO2 reductions of '~3.39 metric tons per cluster annually,' while Section V.F and Table VII state 3.99 metric tons per cluster. These numbers are inconsistent and should be reconciled.","section":"Abstract"},{"comment":"Table VII lists Total Savings (1 Yr, Min) as $1,381, while the text in Section V.F says the combined range is '$1,380 to $2,047.' The $1 versus $0 difference is a minor arithmetic inconsistency.","section":"V.F"},{"comment":"The text writes '0.8 88 MWh monthly' with a stray space; Table VII correctly shows 0.88 MWh. Also, 'approximately 0.8 8 MWh monthly' should be '0.88 MWh.'","section":"V.E"},{"comment":"Reference [37] is titled 'K-Tahp: A Kubernetes Load Balancing Strategy base on TOPSIS+AHP'; the acronym should likely be 'K-TAHP' and the word 'based' is misspelled. In addition, the open-source link in reference [39] points to 'aeris-lab/GreenCube,' which does not match the paper's project name 'GreenPod'; the reproducibility claim should be verified.","section":"II.B"}],"recommendation":"reject","confidential_remarks":"The central empirical comparison may have value if reframed strictly as scheduling-decision energy, but the paper's headline, conclusion, and impact analysis rely on a conflation that inflates the reported savings by roughly two orders of magnitude. In my view, this is a load-bearing error that cannot be fixed by minor edits; it requires either new experiments that measure total workload energy or a complete reframing of the contribution and removal of the extrapolated environmental/economic claims. Given the journal context, I cannot recommend acceptance or even major revision as-is. I would not be opposed to seeing a thoroughly revised manuscript that addresses the measurement methodology and the unit mismatch."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe thing to know about this paper is that the central empirical result is probably real but the headline is not. GreenPod is a TOPSIS-based Kubernetes scheduler, and the authors measure the energy consumed by scheduling decisions across three competition levels and four weighting schemes on a heterogeneous GKE cluster. The measured 39.1% improvement for the energy-centric weighting at medium competition is a reduction in scheduling-decision energy, not a reduction in total workload or cluster energy. The paper itself says so in Table IV: 'Energy Consumed (kJ) quantifies the efficiency of scheduling decisions.' The abstract and conclusion drop that qualification.\n\nWhat is new: TOPSIS for Kubernetes scheduling is not new (KCSS, K-Tahp), but the systematic comparison of weighting schemes on a heterogeneous cluster with different competition levels is a legitimate extension. The experimental setup is reasonable, and the direction of the result makes sense: if you weight energy heavily, you save scheduling energy.\n\nThe soft spot is load-bearing. Section V.E takes the average 19.38% optimization, which applies to scheduling-decision energy of about 0.5 kJ per decision, and multiplies it by an estimate of total job energy (0.024 kWh/job) from a blade-server power model. That calculation assumes the scheduling-energy improvement transfers to total job energy. It does not. The actual scheduling-energy saving is about 0.000025 kWh per decision, roughly 0.1% of the per-job estimate. The annual MWh, CO2, and cost figures are inflated by about two orders of magnitude. There is also a smaller inconsistency: the abstract says 3.39 metric tons of CO2, the body says 3.99.\n\nThe paper lacks error bars or any statistical analysis, and the TOPSIS implementation details are thin (how exactly are weights set? how is energy measured?). The claimed open-source artifact is not verifiable from the paper.\n\nWho is this for? Readers interested in energy-aware Kubernetes scheduling might learn from the experimental design, but the results as presented are not trustworthy. The narrow claim, that a TOPSIS scheduler can reduce scheduling energy, could be a valid small contribution if properly qualified.\n\nRecommendation: send to peer review, because the topic is relevant and the empirical setup is real, but the authors need to fix the extrapolation, qualify the claims, and provide implementation details. As it stands, a reject is defensible; the paper is not publishable without major revision.","headline":"A plausible narrow result about scheduling-decision energy is inflated into cluster-wide savings by a unit mismatch in the extrapolation; worth a look but needs major revision.","tokens_in":11472,"tokens_out":2593,"would_cite":false,"duration_ms":30780,"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":"A TOPSIS-based Kubernetes scheduler cuts scheduling-energy use by up to 39.1%.","keywords":["Kubernetes","TOPSIS","energy-aware scheduling","AIoT","container orchestration","pod placement","multi-criteria decision analysis","edge computing"],"falsifier":"Measure whole-cluster energy (idle, scheduling, and pod execution) over a full day on the same heterogeneous GKE node pool with GreenPod energy-centric scheduling and the default scheduler, using identical workloads; if the total-energy difference is far below the 19.38% used in Section V.E, the annual 10.70 MWh and 3.99-ton CO2 estimates do not follow.","tokens_in":10412,"feed_emoji":"⚡","tokens_out":8217,"duration_ms":82747,"temperature":0.7,"pith_summary":"GreenPod is a custom Kubernetes scheduler that uses TOPSIS, a multi-criteria decision analysis method, to rank nodes on five weighted criteria: execution time, energy consumption, core availability, memory availability, and resource balance. The paper claims that in a heterogeneous Google Kubernetes Engine cluster this approach reduces the energy consumed by scheduling decisions by up to 39.1% relative to the default Kubernetes scheduler, with the largest savings under an energy-centric weighting profile and medium-complexity workloads. The motivation is that the default scheduler optimizes only basic resource availability, which is inadequate for energy-sensitive AIoT workloads spanning cloud and edge. The paper also extrapolates the measured savings to annual energy, CO2, and cost reductions for a production-scale cluster, concluding that energy-aware scheduling is a viable sustainability lever for container orchestration.","feed_headline":"Scheduling fix cuts Kubernetes scheduler energy use by up to 39.1%","feed_subtitle":"A TOPSIS-based pod placer cuts the energy cost of scheduling decisions, with biggest gains on medium workloads.","key_machinery":"The central object is TOPSIS (Technique for Order Preference by Similarity to Ideal Solution), a multi-criteria decision analysis method that ranks each candidate node by its distances to an ideal positive solution and an ideal negative solution, then orders nodes by the resulting closeness coefficient. GreenPod wraps this scoring engine in a scheduling pipeline that profiles workload energy, adaptively weights the five criteria, normalizes the decision matrix, and binds the pod to the top-ranked node through the Kubernetes API. This machinery carries the argument by collapsing five competing objectives into a single node ranking that the scheduler can optimize directly.","core_discovery":"GreenPod replaces the default kube-scheduler's single-objective scoring with a TOPSIS ranking over five weighted criteria (execution time, energy consumption, core availability, memory availability, and resource balance), selecting the node with the highest closeness to an ideal solution. In a heterogeneous Google Kubernetes Engine cluster, the energy-centric weighting profile reduced scheduling-decision energy by up to 39.1% compared with the default Kubernetes scheduler, with the largest gains on medium-complexity workloads and under medium competition. The paper further extrapolates these savings to annual cluster-level energy, CO2, and cost reductions using SURF Lisa job statistics and a blade-server power model.","pith_inferences":["Beyond the paper's measurements, the headline 39.1% figure describes the energy consumed by the scheduling process itself; whether that translates to total cluster energy (including pod execution) is a separate experiment the paper does not report.","Because the resource-efficient profile collapses to 4.86% optimization under high competition, an adaptive scheduler that switches weighting profiles as contention rises could preserve savings where any single fixed profile degrades.","The same TOPSIS pipeline could be tested on GPU-bound or memory-bound AIoT services rather than batch-style linear-regression tasks to see whether the 39% ceiling generalizes to other workload shapes."],"forward_implications":["With energy-centric weighting, GreenPod reports 37.96%, 39.13%, and 33.82% energy optimization over the default scheduler at low, medium, and high competition levels.","Medium competition is the sweet spot: the average optimization across profiles is 24.03%, compared with 18.98% at low and 15.12% at high competition.","Performance-centric weighting produces the smallest energy gains (2.22% to 8.29%), so prioritizing execution speed alone undercuts the sustainability objective.","If the Section V.E extrapolation holds, one SURF Lisa-scale cluster would save about 10.70 MWh and 3.99 metric tons of CO2 per year, and about $1,380 in direct electricity costs.","Energy-centric scheduling allocates work preferentially to energy-efficient node category A, which is why medium-complexity workloads benefit most while light workloads show variable results due to scheduling overhead."],"supporting_citations":[{"why":"Supplies evidence that default Kubernetes scheduling is inadequate for containerized IoT applications in edge environments, the gap GreenPod targets.","marker":"[16]"},{"why":"Shows the native Kubernetes scheduler lacks network and multi-criteria awareness, motivating the weighted TOPSIS approach.","marker":"[17]"},{"why":"Presents KCSS, a prior TOPSIS-based Kubernetes container scheduling strategy that GreenPod extends with a fuller analysis of weighting schemes and workload diversity.","marker":"[18]"},{"why":"Provides the SURF Lisa SLURM job statistics (6,304 jobs per day, workload composition) used in the real-world extrapolation.","marker":"[31]"},{"why":"Supplies the blade-server power model and typical workload parameters used to estimate 0.024 kWh per job.","marker":"[32]"},{"why":"Provides the eGRID CO2 emission factor used to convert annual energy savings into metric tons of CO2.","marker":"[33]"},{"why":"Supplies the commercial electricity rate ($0.1289/kWh) used to calculate annual cost savings.","marker":"[35]"},{"why":"Provides the carbon credit price range ($0.46 to $167 per metric ton) used to estimate the value of avoided emissions.","marker":"[36]"}],"fun_headline_variants":["GreenPod scheduler cuts Kubernetes energy use 39.1%","TOPSIS-based K8s scheduling saves up to 39.1% energy","AIoT scheduler: 39.1% energy savings over default K8s","GreenPod: TOPSIS scheduler trims Kubernetes energy 39.1%","Kubernetes scheduling gets 39.1% more energy-efficient"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The broad annual savings claims depend on applying the 19.38% average optimization, which was measured on scheduling-decision energy, to the 0.024 kWh per-job total-energy estimate, and the paper does not show that scheduling overhead scales to total workload energy.","fun_headline_variants_meta":{"raw":{"variants":["GreenPod scheduler cuts Kubernetes energy use 39.1%","TOPSIS-based K8s scheduling saves up to 39.1% energy","AIoT scheduler: 39.1% energy savings over default K8s","GreenPod: TOPSIS scheduler trims Kubernetes energy 39.1%","Kubernetes scheduling gets 39.1% more energy-efficient"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000138,"raw_usage":{"total_tokens":1072,"prompt_tokens":779,"completion_tokens":293,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":395,"completion_tokens_details":{"reasoning_tokens":193}},"tokens_in":395,"tokens_out":293,"duration_ms":3540,"temperature":1.0,"reasoning_tokens":193,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:30:56.553173+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure whole-cluster energy (idle, scheduling, and pod execution) over a full day on the same heterogeneous GKE node pool with GreenPod energy-centric scheduling and the default scheduler, using identical workloads; if the total-energy difference is far below the 19.38% used in Section V.E, the annual 10.70 MWh and 3.99-ton CO2 estimates do not follow.","supporting_citations":[{"cited_title":"EIS: Edge Information - Aware Scheduler for Containerized IoT Applications","cited_arxiv_id":null,"evidence_quote":"Supplies evidence that default Kubernetes scheduling is inadequate for containerized IoT applications in edge environments, the gap GreenPod targets."},{"cited_title":"Geo -distributed efficient deployment of containers with kubernetes","cited_arxiv_id":null,"evidence_quote":"Shows the native Kubernetes scheduler lacks network and multi-criteria awareness, motivating the weighted TOPSIS approach."},{"cited_title":"KCSS: Kubernetes container scheduling strategy","cited_arxiv_id":null,"evidence_quote":"Presents KCSS, a prior TOPSIS-based Kubernetes container scheduling strategy that GreenPod extends with a fuller analysis of weighting schemes and workload diversity."},{"cited_title":"How do ML jobs fail in datacenters? Analysis of a long -term dataset from an HPC cluster","cited_arxiv_id":null,"evidence_quote":"Provides the SURF Lisa SLURM job statistics (6,304 jobs per day, workload composition) used in the real-world extrapolation."},{"cited_title":"Data center energy consumption modeling: A survey","cited_arxiv_id":null,"evidence_quote":"Supplies the blade-server power model and typical workload parameters used to estimate 0.024 kWh per job."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the eGRID CO2 emission factor used to convert annual energy savings into metric tons of CO2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the commercial electricity rate ($0.1289/kWh) used to calculate annual cost savings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the carbon credit price range ($0.46 to $167 per metric ton) used to estimate the value of avoided emissions."}],"review_version":1}