{"id":"226b3191-acf5-4e5d-bd64-e929701dd694","arxiv_id":"2505.23554","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SLIT is a machine-learning-guided evolutionary scheduler that, in simulation, reduces carbon, water, energy cost, and time-to-first-token relative to two existing LLM serving systems on a synthetic workload derived from a real trace.","lead":"This paper proposes SLIT, a scheduling framework that decides where to run large language model queries across cloud datacenters to balance speed, carbon emissions, water use, and energy cost. A general reader might care because LLM inference is increasingly expensive for the environment, and a practical way to trade off these costs would help cloud operators.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 12's evaporative water model is dimensionally incomplete, undermining the 97–99% water reduction claim.","rationale":"The reader's weakest assumption is that the Section 3 datacenter/energy/carbon/water models faithfully represent real behavior; my concern agrees with this but sharpens it into a specific internal modeling defect in Eq. 12 (and a related index issue in Eq. 16). This is load-bearing because the strongest claim—97–99% water reduction—rests directly on that equation. However, the concern does not change the verdict: CONDITIONAL remains appropriate, since the condition 'validate the simulator with real measurements and rerun the comparisons' is exactly what would resolve it. I did not escalate to REJECT because the paper's central contribution is the multi-objective scheduling framework, and the relative improvements could survive a corrected model; the fix is a concrete re-derivation and recalibration. I also credit the paper for using a real LLM trace, providing pseudocode for the optimizer, and framing the single-objective results as part of a Pareto set rather than a single universal solution. The main missing pieces are empirical validation and parameter disclosure, which the conditional verdict already captures.","tokens_in":10022,"tokens_out":5047,"duration_ms":51634,"concrete_test":"Recompute W_E using a standard cooling-tower evaporation model W_E = H_IT / (ρ_w · c_p · ΔT), with location-dependent ΔT values, or using the latent heat of vaporization, and rerun the SLIT-Water experiment and full Pareto front generation. If the selected datacenters shift or the reported 97–99% water reduction versus Helix/Splitwise changes by more than 5 percentage points, the water claim is dominated by the model error rather than by real scheduling gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The claimed water savings depend critically on the water model in Section 3.3. Eq. 12 states W_E,l,t = H_IT,l,t / H_water, with H_water described as the 'heat capacity of water.' Dividing an energy (J) by a specific heat capacity (J/(kg·K)) gives units of kg·K, not liters of evaporated water; a correct evaporative loss expression needs the temperature rise across the cooling tower (or the latent heat of vaporization). Eq. 13 then defines blowdown as W_B = W_E/(1−D), a nonstandard form that diverges as D→1, and no value for D is provided. Because the simulator is the sole evidence for the headline water reductions, and the paper gives no measured water data, code, or parameter tables, the 97–99% water savings may be artifacts of a dimensionally inconsistent model rather than properties of real datacenters. Additionally, Eq. 16 writes C_Grid,l,t = CI_l,t × E_tot,t without the location index on E_tot, indicating a likely index/unit error in the carbon model. These are internal modeling defects, not merely calibration gaps.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SLIT, a metaheuristic framework that combines a gradient-boosting ML-guided local search with an evolutionary algorithm to schedule LLM inference requests across geo-distributed datacenters. The scheduling objectives are time-to-first-token (TTFT), carbon emissions, water usage, and energy cost. The authors construct synthetic workloads from the BurstGPT trace, model datacenter energy, cooling, water, and carbon behavior in a Python simulator, and compare SLIT-generated plans against Helix and Splitwise. They report very large reductions in carbon, water, energy cost, and TTFT for the single-objective SLIT variants, and a balanced plan that outperforms Helix on all metrics and Splitwise on three of four.","tokens_in":10195,"tokens_out":3978,"duration_ms":39740,"significance":"If the simulation models are trustworthy, the paper addresses a timely and important problem: operational-phase sustainability of LLM inference. The authors should be credited for building on a real LLM request trace, for providing algorithm pseudocode, and for comparing against two published serving systems rather than only toy baselines. The Pareto-front presentation is a useful practical feature. However, all headline numerical claims rest on a simulator whose physical models contain a dimensional error and whose 'validation' is asserted without evidence. The reported 97-99% reductions are therefore not yet evidence of real-world gains.","major_comments":[{"comment":"The evaporative water model is dimensionally incorrect. Eq. (12) defines W_E,l,t = H_IT,l,t / H_water, where H_water is described as the 'heat capacity of water.' Dividing an energy (J) by a specific heat capacity (J/(kg·K)) yields kg·K, not liters or kilograms of evaporated water. A physically meaningful expression requires the latent heat of vaporization (or the temperature rise across the cooling tower). Since the water savings are a headline result, this error is load-bearing. Relatedly, Eq. (13) defines blowdown as W_B = W_E/(1-D), which diverges as D approaches 1, and no value or source for D is ever given. The water total in Eq. (15) thus depends critically on an unspecified, nonstandard parameter. The water-reduction claims in Section 6 and Fig. 4 should be recomputed with a corrected model and a stated D value, and the sensitivity to D should be reported.","section":"Section 3.3, Eqs. (12)-(13)"},{"comment":"The carbon model contains internal index and notation errors. Eq. (16) writes C_Grid,l,t = CI_l,t × E_tot,t, but E_tot,t has no location index even though the left-hand side is location-specific; it should be E_tot,l,t. Eq. (17) uses W_G,l,t, but the text and Eq. (14) define W_Grid,l,t, not W_G,l,t. These are not merely cosmetic typos because the carbon objective C_tot,t in Eq. (18) is defined as the sum of C_Grid and C_W; the mis-indexing changes the computed carbon footprint whenever datacenters differ in size or grid carbon intensity. The carbon-reduction percentages in Section 6 must be verified after correcting these expressions.","section":"Section 3.4, Eqs. (16)-(17)"},{"comment":"The paper states in Section 6, 'We developed and validated a Python-based simulator,' but no comparison to measured energy, water, or carbon data is presented anywhere. The simulator is the sole source of evidence for the claimed reductions. In addition, the experiments are single-run: no error bars, no multiple independent trials, and no sensitivity analysis for key parameters (CoP, D, water intensity, carbon intensity, time-of-use prices, node TDP). Because SLIT's search is stochastic (random initialization, random parent selection, mutation), single-run results cannot support the reported precise percentages. The authors should report multiple seeds with confidence intervals and at least a calibration of the energy and water models against published datacenter measurements or a validated reference model.","section":"Section 6, simulator validation and experimental design"},{"comment":"Helix and Splitwise are not designed to optimize carbon, water, or energy cost; they optimize serving throughput and latency under heterogeneous-GPU constraints. Comparing a multi-objective sustainability scheduler against them on sustainability metrics therefore makes large reductions almost inevitable and does not by itself demonstrate that SLIT is a better multi-objective optimizer. Also, the paper repeatedly calls the output a 'Pareto optimal' solution set (e.g., Section 1 and Section 6), but a metaheuristic with dominance-based population updates provides no Pareto-optimality guarantee. The authors should either compare against a sustainability-aware baseline or report the baselines' performance on their own objectives, and they should present multi-objective quality indicators (e.g., hypervolume) computed on multiple runs, or restrict the claim to an approximate Pareto set.","section":"Section 6, Fig. 4 and comparison baselines"}],"minor_comments":[{"comment":"There are two subsections numbered 3.3 ('Energy Cost Model' and 'Water Model'); the subsequent numbering (3.4 Carbon Model) is consequently misleading.","section":"Section 3, headings"},{"comment":"Reference [27] contains a typo, 'Enviornment' should be 'Environment.'","section":"References"},{"comment":"No table of parameter values is provided for the simulator (e.g., CoP_l, D, WI_l,t, CI_l,t, TOU_l,t, bandwidth, node counts, GPU memory capacities). Without these values, the experiments cannot be reproduced or scrutinized.","section":"Section 6, reproducibility"},{"comment":"Line 22 of the pseudocode says 'If prediction != observed do' with no specification of how the fallback default plan is generated or how many requests are affected when the predictor is wrong; the impact of prediction error on the reported metrics is not quantified anywhere.","section":"Fig. 3, Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The central contribution is conditional on the simulator's fidelity, and the dimensional error in Eq. (12) plus the unsubstantiated 'validated' claim in Section 6 are serious. I recommend major revision rather than rejection because the errors appear local and fixable in principle: the water and carbon models can be corrected, parameters can be supplied, and the experiments can be rerun with multiple trials and proper baselines. If the authors do not provide those corrections and validation, the headline reduction claims should not be considered established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hey [Name],\n\nQuick read on 2505.23554 for you. The paper proposes SLIT, a scheduler for LLM inference across geo-distributed datacenters that simultaneously optimizes time-to-first-token, carbon, water, and energy cost. That four-way co-optimization with a Pareto front is genuinely new relative to the literature they cite—Helix and Splitwise are throughput/latency focused, and prior sustainable-cloud work doesn't handle LLM inference. The problem is timely and the combination of gradient-boosting-guided local search with an evolutionary algorithm is a sensible metaheuristic choice. So there is a real contribution here, in scope and in framing.\n\nWhat the paper does well: the workload model uses the BurstGPT trace, they model KV cache memory, migration latency, and per-node power states, and they present per-epoch comparisons alongside the aggregate numbers. The Pareto framing for letting an operator pick a trade-off is practical.\n\nBut the soft spots are serious. The simulator is the only evidence, and the text says it is 'validated' without showing a single comparison to measured energy, water, or carbon data. There are no error bars, no multiple independent trials, and no code or parameter tables. The baselines Helix and Splitwise do not optimize any sustainability objective, so the 97-99% reductions are largely an artifact of comparing a sustainability-optimizing optimizer against sustainability-blind baselines; a sustainability-aware baseline like GreenCourier or the Qi et al. work would be the honest comparison.\n\nWorse, the water model in Section 3.3 does not survive dimensional analysis. Equation 12 divides an energy by a heat capacity, which gives kg·K, not liters of evaporated water. The correct form needs a latent heat or a temperature rise across the cooling tower. Equation 13 blows up as D→1, and D is never given a value. Equation 16 has a missing location index on E_tot. These are internal modeling defects, not just calibration gaps. The headline water savings rest on this model, so the 97-99% water claim is untrustworthy as written.\n\nThat said, the core idea is salvageable. The carbon and cost models look more conventional, and the metaheuristic is reasonable. If the authors fix the water model, validate the simulator against real datacenter data (or at least publish the code and parameter values), and include sustainability-aware baselines, this could be a solid systems paper. In its current state, however, the results overstate what is actually shown.\n\nWho's the audience? Researchers in sustainable cloud computing and LLM serving who want to see how one might combine these four objectives. I'd send it back for major revision, not desk reject—the problem space is important enough and the framework is novel enough to deserve referee time. But I would not cite the water numbers until they're redone.\n\nBest,\n[Your name]","headline":"A genuinely new four-objective LLM scheduling framework whose headline water savings rest on a dimensionally wrong model and an unvalidated simulator; worth revising but not citing yet.","tokens_in":10764,"tokens_out":3332,"would_cite":false,"duration_ms":28583,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a new scheduling framework, SLIT, co-optimizes time-to-first-token, carbon emissions, water use, and energy cost for LLM inference across geo-distributed datacenters, and in simulation reduces each by 73-99% versus…","keywords":["carbon-aware scheduling","water-efficient datacenters","LLM inference","geo-distributed clouds","multi-objective optimization","Pareto front","evolutionary algorithm","gradient boosting"],"falsifier":"Run SLIT's recommended plans on a small real geo-distributed testbed for 24 hours, recording actual energy, water, carbon, and TTFT, and compare with simulator predictions; large discrepancies in cooling or water overhead would show the reductions are simulation artifacts.","tokens_in":9806,"feed_emoji":"🌱","tokens_out":5612,"duration_ms":47586,"temperature":0.7,"pith_summary":"This paper tries to establish that the environmental cost of serving LLM inference requests can be cut substantially by smart request scheduling across geo-distributed cloud datacenters. It introduces SLIT, a framework that co-optimizes four objectives at once: time-to-first-token (TTFT), carbon emissions, water usage, and energy cost, returning a Pareto set of scheduling plans rather than a single answer. In simulation, the single-objective SLIT variants reduce carbon emissions by 98-99%, water usage by 97-99%, energy cost by 96-99%, and TTFT by 73-81% compared with the Helix and Splitwise baselines. A balanced variant, SLIT-Balance, beats Helix on all four metrics and Splitwise on three while keeping TTFT competitive. If the underlying datacenter models are faithful, the paper's approach would let LLM providers shrink the inference-phase footprint that now dominates LLM costs.","feed_headline":"LLM scheduling slashes inference carbon, water, and cost by up to 99%","feed_subtitle":"SLIT's Pareto-optimal plans beat Helix and Splitwise on carbon, water, and cost with competitive TTFT.","key_machinery":"The mechanism that carries the argument is the SLIT metaheuristic, iterating ML-guided local search with an evolutionary algorithm. The search begins from a partially random population seeded with two extreme plans (even distribution and single-location scheduling); each plan is locally searched, and the trajectories are used to retrain a gradient-boosting predictor every $freq$ iterations. An evolutionary loop then randomly selects parent plans from the current nondominated population, applies crossover and mutation to generate children, and keeps only dominant plans in the population, yielding a real-time Pareto front. Every plan is scored against the paper's datacenter model: IT energy from power states (Eq. 5), cooling energy with $E_{cooling}=3\\times E_{CRAC}$ (Eq. 8), a 13% power-conditioning overhead (Eq. 9), water from evaporative, blowdown, and grid components (Eqs. 12-15), and carbon from grid intensity plus water-treatment emissions (Eqs. 16-18). The workload predictor (a set of incrementally trained linear regression models) supplies the request count for the upcoming 15-minute epoch.","core_discovery":"The central claim is that scheduling LLM inference requests is the key lever for sustainability: by choosing which of several global datacenters (with heterogeneous A100/H100 nodes, different grid carbon intensities, water intensities, time-of-use electricity prices, and cooling costs) handles each request, a provider can simultaneously cut carbon, water, and energy cost without sacrificing perceived latency. SLIT finds these plans by combining an ML-guided local search, in which a gradient-boosting model trained on previous search trajectories suggests promising starting points, with an evolutionary algorithm that crosses and mutates existing good plans to escape local optima. Over a 24-hour scaled workload derived from the BurstGPT trace, the paper reports that SLIT's Pareto-optimal solutions dominate the Helix mixed-integer-programming scheduler and the Splitwise phase-splitting queue scheduler on the co-optimization objectives, with per-epoch plots showing the balanced SLIT solution staying near Splitwise's TTFT while emitting far less carbon and using far less water.","pith_inferences":["The magnitude of the claimed reductions depends entirely on the Section 3 model constants; if real cooling energy is not three times CRAC energy or the blowdown ratio differs, the absolute percentages would shift even if the relative ranking of schedulers does not.","A natural extension, not explored in the paper, is feeding renewable-energy and carbon-intensity forecasts into the same metaheuristic to shift load toward cleaner hours, not just cleaner regions.","The Pareto-front formulation could be exposed as an operator-facing policy knob, letting a provider set a maximum carbon budget per request and have SLIT pick the best latency within that budget.","A real 24-hour testbed deployment with meter readings would be the decisive next check, converting the simulation's 97-99% claims into an operational claim."],"forward_implications":["Cloud operators can select, from the Pareto front SLIT returns, a plan that minimizes any single objective (carbon, water, cost, or TTFT) or a balanced compromise, within the 15-minute decision window.","SLIT-Balance outperforms Helix on all four objectives and Splitwise on carbon, water, and cost while keeping TTFT competitive, showing sustainability does not have to trade away responsiveness.","The framework scales to realistic conditions: 12 global datacenters, 6 heterogeneous node types, 1000 nodes per site, and a workload scaled from the real BurstGPT trace with 0.5x request delays, 3x token counts, and 10x request counts.","The workload predictor lets the scheduler react to the rapid request-intensity changes observed in real LLM traces, avoiding stale plans for the next epoch."],"supporting_citations":[{"why":"Helix, the mixed-integer linear programming LLM scheduler that serves as the primary performance and sustainability baseline.","marker":"[16]"},{"why":"Splitwise, the phase-splitting queue scheduler that serves as the second baseline.","marker":"[17]"},{"why":"BurstGPT, the real-world ChatGPT request trace whose trends and scaled workload drive all experiments.","marker":"[19]"},{"why":"Source for the mechanical cooling and CRAC energy relations used in Eq. 8.","marker":"[23]"},{"why":"Source for the 13% power-conditioning overhead in Eq. 9.","marker":"[24]"},{"why":"Source for water intensity values per electricity source used in Eq. 14.","marker":"[25]"},{"why":"Source of the regression-based workload predictor method used in Section 5.1.","marker":"[28]"},{"why":"Gradient boosting machine used for the ML-guided local search.","marker":"[29]"}],"fun_headline_variants":["SLIT scheduler cuts LLM carbon, water, cost by up to 99%","Geo-distributed LLM scheduling slashes environmental impact","SLIT co-optimizes LLM serving across clouds for sustainability","New scheduler trims LLM footprint without slowing responses","Water-wise, carbon-cheap LLM inference via smart scheduling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the paper's equations for datacenter energy, cooling, water, and carbon (with cooling set to three times CRAC energy and a 13% power-conditioning overhead) faithfully describe how real datacenters behave, since the simulator is never checked against measured data.","fun_headline_variants_meta":{"raw":{"variants":["SLIT scheduler cuts LLM carbon, water, cost by up to 99%","Geo-distributed LLM scheduling slashes environmental impact","SLIT co-optimizes LLM serving across clouds for sustainability","New scheduler trims LLM footprint without slowing responses","Water-wise, carbon-cheap LLM inference via smart scheduling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000654,"raw_usage":{"total_tokens":3011,"prompt_tokens":973,"completion_tokens":2038,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":1950}},"tokens_in":589,"tokens_out":2038,"duration_ms":14950,"temperature":1.0,"reasoning_tokens":1950,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:43:01.470502+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SLIT's recommended plans on a small real geo-distributed testbed for 24 hours, recording actual energy, water, carbon, and TTFT, and compare with simulator predictions; large discrepancies in cooling or water overhead would show the reductions are simulation artifacts.","supporting_citations":[{"cited_title":"Helix: Distributed Serving of Large Language Models via Max - Flow on Heterogenous GPUs,","cited_arxiv_id":null,"evidence_quote":"Helix, the mixed-integer linear programming LLM scheduler that serves as the primary performance and sustainability baseline."},{"cited_title":"Splitwise: Efficient Generative LLM Inference Using Phase Splitting,","cited_arxiv_id":null,"evidence_quote":"Splitwise, the phase-splitting queue scheduler that serves as the second baseline."},{"cited_title":"BurstGPT: A Real-world Workload Dataset to Optimize LLM Serving Systems,","cited_arxiv_id":null,"evidence_quote":"BurstGPT, the real-world ChatGPT request trace whose trends and scaled workload drive all experiments."},{"cited_title":"A survey on data center cooling systems: Technology, power consumption modeling and control strategy optimization,","cited_arxiv_id":null,"evidence_quote":"Source for the mechanical cooling and CRAC energy relations used in Eq. 8."},{"cited_title":"A Review of Data Centers Energy Consumption and Reliability Modeling,","cited_arxiv_id":null,"evidence_quote":"Source for the 13% power-conditioning overhead in Eq. 9."},{"cited_title":"Water Use of Electricity Technologies: A Global Meta -Analysis,","cited_arxiv_id":null,"evidence_quote":"Source for water intensity values per electricity source used in Eq. 14."},{"cited_title":"Mu: An Efficient, Fair and Responsive Serverless Framework for Resource-constrained Edge Clouds,","cited_arxiv_id":null,"evidence_quote":"Source of the regression-based workload predictor method used in Section 5.1."},{"cited_title":"Greedy Function Approximation: A Gradient Boosting Machine,","cited_arxiv_id":null,"evidence_quote":"Gradient boosting machine used for the ML-guided local search."}],"review_version":1}