Policy-aware GPU resource allocation for national supercomputing.

Policy-aware GPU resource allocation for national supercomputing - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Mar 6;16:12438. doi: 10.1038/s41598-026-42625-6 Search in PMC Search in PubMed View in NLM Catalog Add to search Policy-aware GPU resource allocation for national supercomputing Hyungwook Shim Hyungwook Shim 1 Korea Institute of Science and Technology Information, Daejeon, Korea Find articles by Hyungwook Shim 1, ✉ Author information Article notes Copyright and License information 1 Korea Institute of Science and Technology Information, Daejeon, Korea ✉ Corresponding author. Received 2025 Sep 1; Accepted 2026 Feb 26; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13083863  PMID: 41792216 Abstract The rapid expansion of AI, large-scale simulation, and data-intensive research has increased demand for GPU resources in national supercomputing systems. However, allocation mechanisms in many centers remain primarily workload-driven, which may not fully reflect evolving policy priorities across scientific domains. This study proposes a policy-integrated optimization framework that combines a target-aware static estimator with a dynamic runtime reallocation controller to incorporate policy objectives into operational allocation decisions. The framework internalizes a policy target vector , constructed from stated priorities and historical usage, within both estimation and control stages. Using empirical demand curves under a rolling out-of-sample protocol, the proposed approach reduces policy-alignment error relative to uncontrolled and demand-driven baselines (MAE from 8.03 to 1.30% and RMSE from 9.59 to 1.66%), while maintaining GPU utilization above 92% and comparable throughput and queueing performance. Sensitivity analyses indicate stable behavior across tested parameter ranges. These findings suggest that policy-aware resource allocation can be implemented within existing scheduling environments without materially affecting core operational metrics, providing a structured approach to aligning supercomputing governance with strategic priorities. Supplementary Information The online version contains supplementary material available at 10.1038/s41598-026-42625-6. Keywords: GPU resource allocation, Supercomputing policy, Dynamic resource allocation algorithm, Artificial intelligence, Science and technology strategy Subject terms: Engineering, Mathematics and computing Introduction The rapid advancement of artificial intelligence, large-scale scientific simulations, and data-intensive research has dramatically increased the demand for GPU resources. National supercomputing infrastructures have evolved into strategic assets that underpin technological competitiveness and scientific innovation. In response, countries have accelerated infrastructure investments to secure leadership in critical technology domains 1 . Despite these trends, most supercomputing centers continue to rely on demand-oriented allocation schemes, where resources are distributed largely based on average runtime and job length. For example, NERSC in the United States, RIKEN in Japan, and EuroHPC Joint Undertaking in Europe employ scheduling and allocation rules that prioritize utilization stability and long-running workloads 2 – 4 . This operational paradigm systematically channels capacity toward high-load domains, inadvertently under-provisioning fields of strategic national importance. As resource competition intensifies, equitable allocation that reflects policy priorities is increasingly recognized as a key principle of sustainable supercomputing governance. Ensuring balanced development requires a resource allocation framework that aligns operational decisions with policy-defined strategic domains—for example, foundational research, emerging technologies, or nationally prioritized R&D fields. Without such alignment, purely demand-driven allocation risks reinforcing structural inequities and weakening strategic competitiveness. In parallel, global supercomputing expansion faces practical constraints from energy, environmental, and economic pressures. This intensifies the need to optimize not only efficiency but also policy coherence and public value. Addressing this challenge requires moving beyond conventional fairness- or SLA-based scheduling approaches and embedding policy objectives directly within allocation mechanisms. Our study contributes to this emerging direction by introducing a policy-integrated optimization framework that: Internalizes a policy target vector in the optimization objective to formally align policy and operations; Integrates a two-stage design consisting of a static, target-aligned estimator and a dynamic controller for real-time reallocation; Provides interpretable trade-off parameters ( ) that balance structural similarity with demand-driven signals. This approach bridges the gap between high-level policy priorities and operational scheduling, offering a scalable mechanism to achieve fairness and efficiency simultaneously. Literature review The surge in demand for GPU resources driven by the rapid expansion of AI workloads, large-scale simulations, and data-centric research has transformed resource allocation in supercomputing environments into a multi-objective optimization problem. Beyond technical performance, modern allocation strategies are increasingly expected to ensure fairness, energy sustainability, and alignment with policy priorities. Existing work (Table 1 ) in this area can be broadly divided into two strands: technical approaches that focus on optimization and scheduling efficiency, and policy-driven approaches that emphasize strategic and institutional priorities. Table 1. Classification of prior work on GPU resource allocation. Category Representative work/program Core contribution Technical Mo (2024) 5 OEF framework for balancing fairness and throughput Zeng (2021) 6 Federated learning optimization for energy savings Ramamoorthi (2021) 7 AI-driven predictive control for dynamic cloud loads Tai (2023) 8 Green-IT based time-aware scheduling Wang (2025) 9 MDP-based scheduling (AdaGap) DRF/SLA schedulers Exogenous utility-based fairness and service guarantees Policy-driven INCITE, ALCC (DOE, U.S.) Mission-oriented allocation NSF ACCESS (U.S.) Balanced peer-review-based allocation across fields and institutions EuroHPC, PRACE, DECI (EU) Evaluation of scientific impact and exploitation plans EPSRC/ARCHER2 RAP (U.K.) Alignment with national strategic research domains Open in a new tab Comparative remarks are based on the degree of policy integration and runtime adaptability derived from public allocation guidelines. Technical approaches to resource optimization Technical approaches primarily aim to improve the operational efficiency, throughput, and adaptability of GPU resource allocation, often under heterogeneous and time-varying demand. Mo (2024) proposes the OEF (Optimal Efficiency and Fairness) framework to reconcile the intrinsic tension between fairness and throughput in heterogeneous GPU clusters. By framing scheduling as a global optimization problem, OEF improves throughput by up to 32% over state-of-the-art baselines while guaranteeing fairness in both cooperative and non-cooperative environments 5 . Zeng (2021) addresses energy optimization in federated learning settings through joint optimization of bandwidth allocation, CPU–GPU partitioning, and speed scaling to minimize energy consumption at fixed accuracy levels 6 . Ramamoorthi (2021) introduces an AI-driven predictive control system that dynamically adapts to variable cloud loads, while Tai (2023) incorporates Green-IT principles to design time-aware, energy-constrained scheduling mechanisms 7 , 8 . Wang (2025) formulates the resource allocation problem as a Markov Decision Process (MDP) via AdaGap, demonstrating latency and gap reductions in heterogeneous clusters using production traces 9 . Recent studies have further advanced GPU scheduling and fairness-aware optimization in large-scale or multi-cluster environments. For example, Wang (2025) introduces a deterministic scheduling framework that integrates user collaboration to improve workload predictability and fairness in deep learning clusters 10 . Wu (2025) addresses cross-cluster load balancing and resource fragmentation through adaptive multi-cluster coordination 11 . In parallel, Huang (2025) provides a comprehensive taxonomy of emerging trends in accelerator resource management, highlighting the growing need for secure and policy-aware allocation mechanisms 12 . These works collectively underscore the importance of aligning fairness, efficiency, and scalability—issues also central to the framework proposed in this study. Classic fairness-based (DRF) and SLA-aware schedulers similarly encode exogenous utilities, caps, or weights to maintain equity and service guarantees. However, these methods treat policy objectives as external constraints rather than endogenous optimization targets. As a result, while they excel in operational performance, they remain limited in addressing policy alignment or strategic domain balancing. Policy-driven approaches to resource allocation Policy-driven approaches emphasize alignment with strategic domains, scientific impact, and societal or national priorities. These mechanisms are typically embedded in national and regional HPC allocation frameworks. In the United States, the INCITE and ALCC programs of the Department of Energy allocate large-scale computational resources to projects that contribute to mission areas such as energy and security. The NSF ACCESS program (formerly XSEDE) evaluates applications through peer review panels that balance field and institutional representation, aiming to ensure a fair distribution across research communities. In Europe, EuroHPC Regular Access and the PRACE/DECI programs emphasize the advancement of the European HPC ecosystem and balanced distribution, evaluating scientific merit, impact, and exploitation plans. In the United Kingdom, EPSRC/ARCHER2 RAP explicitly prioritizes alignment with nationally strategic research domains 10 – 14 . While these programmatic systems successfully embed policy priorities through domain quotas, caps, and review criteria, they are mostly ex-ante and coarse-grained. Policy objectives are not formalized as explicit vectors for optimization, nor are they tightly coupled to runtime control mechanisms capable of reducing deviations under time-varying demand. This structural disconnect between policy design and operational execution remains a significant gap in current practice. Identified gaps and this study’s contribution Across both strands of research, two critical gaps emerge: Lack of policy internalization: Policy goals are typically treated as external constraints rather than endogenous optimization objects. Weak coupling between design and operations: While programmatic allocation policies establish strategic priorities, they lack systematic runtime mechanisms to drive resource distribution toward those priorities as demand evolves. This study addresses these gaps by introducing a policy-integrated framework that (i) embeds a policy target vector directly into the objective function of a static estimator, yielding interpretable trade-off parameters ( ) between structural similarity and demand signals; and (ii) links the estimator to a runtime controller that enforces effective caps, reclaims excess, and reallocates capacity to domains with positive policy shortfall. In doing so, it establishes a principled bridge between policy design and operational scheduling, rigorously evaluated through rolling out-of-sample protocols and stress tests. While previous programs such as INCITE, ACCESS, and EuroHPC incorporate policy priorities into allocation decisions, they vary substantially in operational adaptability and feedback integration. Table 1 summarizes these distinctions, contrasting technical optimization frameworks with policy-driven approaches and their respective degrees of runtime flexibility. Research methodology and design This section presents the methodological architecture of the proposed optimization framework for policy-oriented GPU resource allocation. The core objective is to establish a flexible allocation mechanism that integrates structural similarity among scientific domains with demand-driven usage intensity, thereby aligning operational allocations with strategic policy priorities. The model achieves this by combining cosine similarity with a weighted scoring function and optimizing adjustment coefficients to minimize deviation from policy targets. This architecture is fully simulation-compatible, enabling quantitative performance evaluation under realistic demand curves. Data structure and vector definition The empirical foundation of this study consists of system log data from Neuron, Korea’s national GPU-focused supercomputing infrastructure, and NERSC, a representative U.S. high-performance computing (HPC) system. These two systems provide contrasting allocation regimes: Neuron currently operates without domain-specific allocation weights, whereas NERSC is widely recognized for implementing more balanced, domain-aware allocation policies. Using these two datasets enables comparative vectorization of domain-level GPU usage patterns, which serves as the empirical basis for determining optimized allocation ratios 10 , 11 . For each domain , GPU usage characteristics are represented as a two-dimensional vector comprising average job runtime and the proportion of long-duration jobs exceeding ten hours: 1 : Average job execution time (Wall Time) for domain . : Proportion of long-duration jobs exceeding 10 h in domain . A system-level reference vector is defined for Neuron by aggregating job data over the same horizon: where exponential weighted moving average (EWMA) smoothing is applied. All vector components are min–max normalized to [0,1], and L2 normalization is performed prior to cosine similarity evaluation. Importantly, represents the operational baseline of the system—that is, typical job duration and tolerated long-job share—and is used as a structural anchor, not a policy signal. For Neuron, these metrics were indirectly estimated from operational statistics, such as the distribution of job counts by GPU core intervals and wall time. In contrast, the NERSC job-level logs included detailed per-job GPU usage time, runtime, and scientific domain metadata, enabling more direct calculation. To ensure comparability despite differences in data collection methods, common metrics were selected as shown in Table 2 . Table 2. Vector elements for GPU usage comparison between Korea (Neuron) and the U.S. (NERSC). Variable Description Neuron dataset (Korea) NERSC dataset (U.S.) Average job execution time (Wall Time basis) Estimated using average wall time(hour) Calculated from runtime_sec (converted to hour) Proportion of long jobs (> 10 h) Derived from job time distribution Filtered based on runtime Open in a new tab Design of optimization indicators Two indicators—cosine similarity and weighted score—were designed to capture distinct but complementary aspects of domain-level GPU usage characteristics. Cosine similarity measures structural alignment with the system baseline, while the weighted score reflects absolute usage intensity, enabling policy-oriented allocation. Cosine similarity Cosine similarity quantifies directional alignment between each domain vector and the Neuron reference vector . It is defined as Eq. ( 2 ): 2 Values range from 0 to 1, with higher values indicating greater structural similarity in usage profiles. Because cosine similarity focuses on proportional patterns rather than scale, it is well suited for identifying domains that exhibit usage structures similar to the system baseline. However, this indicator does not account for absolute consumption levels, which motivates the introduction of a complementary weighted score. Weighted score The weighted score reflects absolute usage characteristics and their alignment with policy-relevant factors. Each domain’s average runtime and long-job ratio is min–max normalized to [0,1]. Equal weights α and β are applied to generate a scalar score for each domain (Eq. ( 3 )): 3 This score allows domains with usage characteristics similar to Neuron to be quantitatively distinguished. The final policy-oriented allocation share for each domain was calculated as in Eq. ( 4 ), where denotes the total number of analyzed domains: 4 The characteristics of the two indicators are summarized in Table 3 below. Table 3. Characteristics of indicators. Item Cosine similarity Weighted score (policy allocation share) Definition Similarity of vector directions (proportional patterns) Weighted average of normalized values Purpose Assess similarity of structural usage patterns Evaluate suitability for allocation and policy priorities Mathematical basis Inner product, scale-invariant Scale-based, reflects absolute values Output range 0–1 Proportion (0–1) and relative ranking possible Open in a new tab Optimization model We formalize the static estimator that maps domain descriptors to predicted allocation shares while explicitly aligning them with the policy target vector . The estimator balances two signals: structural similarity to the system’s reference profile and demand intensity observed in the logs. Let index scientific domains. We construct two non-negative, unit-sum vectors: the structural score (cosine-similarity to the Neuron reference) and the demand-based score (scale-oriented). By construction, . The raw prediction combines both indicators as in Eqs. ( 5 ), 5 The normalized predicted share is given by Eqs. ( 6 ), 6 Because of the normalization in (6), scaling ( ) by a common positive constant leaves unchanged; we therefore impose for identifiability and interpretability. Let denote the policy target vector. In this study, T is an empirically derived policy baseline representing the intended proportional distribution of GPU resources across scientific domains. It was constructed from three observable components: (i) the 3-year average share of national R&D investment per domain, (ii) policy-designated priority domains drawn from official government strategy documents, and (iii) historical GPU usage shares from the Neuron system to preserve operational continuity. Each component was normalized and combined with equal weights (1:1:1) to form a balanced target vector. Let , , and denote the normalized vectors corresponding to (i) R&D investment share, (ii) policy-designated priority share, and (iii) historical GPU usage share, respectively. The baseline policy target vector is defined as: This equal-weight construction serves as a neutral aggregation rule that avoids ex ante prioritization among funding intensity, formal policy designation, and operational continuity. Sensitivity to alternative target specifications that emphasize different domain groups is reported in Supplementary Information A.10. The resulting T remained time-invariant during simulation, while workload demand curves evolved dynamically at 4-hour intervals for evaluation. At evaluation time , ( ) are chosen by minimizing the weighted MSE in Eqs. ( 7 ), 7 where is computed from ( 5 ),( 6 ) using features available up to , and the loss weights are constructed only from information up to : ], Here, is the EWMA-smoothed historical usage share of domain accumulated up to . The EWMA smoothing parameter (ρ) controls the stability–responsiveness trade-off in lagged usage estimation. We adopt ρ = 0.3 as a baseline and confirm robustness over alternative values (ρ ∈ {0.1, 0.3, 0.5}); results are reported in Supplementary Information A.7. This construction makes measurable with respect to past information and therefore exogenous to the prediction at , removing the self-referential weighting that arises when appears both in prediction and in the loss. We choose via blocked rolling cross-validation that respects temporal order. For each , the set is partitioned into contiguous folds; parameters are selected by minimizing the average validation loss over those folds, and performance is then reported out-of-sample at using Eq. ( 7 ). We slide across the horizon and aggregate OOS(Out-Of-Sample) MAE/RMSE. Numerical minimization uses scipy.optimize.minimize with the box/simplex constraints , . The parameter emphasizes structural alignment (cosine similarity), while emphasizes demand intensity. The deviation also serves as the state signal for the dynamic controller in Sect.  5, enabling a consistent policy–operation bridge between static estimation and runtime reallocation. To validate robustness, we additionally evaluate model performance under two alternative weighting schemes: Uniform weights : , Variance-based weights: As reported in the Supplementary Information, the main conclusions remain qualitatively consistent across all weighting configurations. Results of resource allocation optimization This study proposed an integrated resource allocation optimization model combining a cosine similarity-based structural similarity metric ( ) and a demand-driven weighted score ( ) that reflects the proportion of long-duration jobs, with the aim of achieving rational GPU resource allocation across scientific domains. The optimization yielded α = 0.1631 and β = 0.8369, indicating that the demand-based metric exerted a more substantial influence on aligning predicted allocations with policy targets. These results demonstrate that the predicted allocation ratios were adjusted to closely approximate the target proportions and were also confirmed by visualization of the loss-minimizing optimal combination on the α/β plane (Fig. 1 ). Notably, compared to allocations derived solely from individual indicators (α or β), the integrated model achieved more balanced control of relative errors across domains, thereby offering a practical solution that simultaneously supports flexibility and legitimacy in policy-driven resource management. Fig. 1. Open in a new tab Optimization outcome. Table 4 summarizes the predicted allocation ratios, target ratios, and the corresponding errors for each domain under the optimal combination of α = 0.1631 and β = 0.8369. For example, Computer Science achieved near-perfect alignment, with a predicted ratio (0.3861) extremely close to the target (0.39), resulting in an almost negligible error (−0.0039). In contrast, Materials showed the most pronounced under-allocation, with a predicted ratio of 0.1642 falling 17.58% points below the target of 0.34. Conversely, Earth Science and Engineering exhibited over-allocations of approximately + 9%p and + 9.8%p, respectively. These outcomes suggest that structural trade-offs emerged between domains prioritized by policy and those characterized by higher demand, indicating that relative sacrifices in certain areas contributed to compensatory adjustments in others. The table underscores that while the , optimization minimized the overall loss function, it did not entirely eliminate absolute errors in individual domains. Table 4. Estimated allocation ratio compared to target allocation. Domain Estimated allocation ratio Target ratio Error Computer science 0.39 0.39 0.00 Materials 0.16 0.34 -0.18 Chemistry 0.12 0.18 -0.06 Fusion 0.11 0.06 0.05 Earth science 0.11 0.02 0.09 Engineering 0.11 0.01 0.10 Open in a new tab In estimating the allocation shares, employing either cosine similarity ( ) or the weighted score ( ) as a single criterion also yielded intervals where the Weighted MSE attained local minima. Specifically, when optimizing only while fixing as , or conversely optimizing only , U-shaped loss curves were observed, with certain points minimizing the deviation between predicted and target allocations. However, relying solely on a single parameter resulted in persistent errors across domains, with tendencies toward significant over- or under-allocation in specific areas. This limitation highlighted the challenge of achieving both policy equity and operational efficiency through a unidimensional approach. By contrast, when and were jointly treated as independent variables and optimized simultaneously, the resulting predictions more precisely converged with the target allocations, and the overall MSE was markedly reduced (Fig. 2 ). This evidence demonstrates that a dual-weighted optimization framework, which integrates both indicators, is practically superior to single-metric strategies for resource allocation policymaking. Fig. 2. Open in a new tab Dual-weighted optimization curve. Dynamic resource allocation simulation As confirmed in the optimization results (Table 3 ), certain domains, such as Materials, exhibited substantial discrepancies between the estimated allocation ratio and the target allocation ratio (e.g., − 17.58% points). This indicates that static weighted optimization alone is insufficient to fully achieve policy-prescribed targets. Addressing such structural discrepancies requires not merely one-time numerical adjustment but the integration of dynamic mechanisms during operation. Accordingly, we propose and simulate a domain-specific control algorithm that adjusts allocations in response to actual requests and utilization. The simulation uses time-series demand curves for six domains over one week at four-hour intervals (28 windows). All results are reported in normalized capacity units: in each window the total capacity is scaled to 1.0, and all requests/allocations are expressed as shares. This renders MAE/RMSE of allocation ratios invariant to absolute capacity. For practical grounding, we calibrate the normalized series to Neuron’s historical weekly delivered GPU-hours (aggregate, anonymized) and verify that trajectories and errors are unchanged (Supplementary Information). Dynamic resource allocation algorithm To ensure robust alignment with policy priorities under variable workload conditions, we implement a dynamic domain-level allocation mechanism. The controller operates on normalized GPU capacity units, where each 4-hour time window is scaled to a total capacity of 1.0, and all domain-level requests and allocations are expressed as proportional shares. This normalization renders evaluation metrics—such as MAE and RMSE of allocation ratios—invariant to changes in absolute capacity. For empirical calibration, the normalized curves are anchored to Neuron’s historical weekly GPU-hour delivery (aggregated and anonymized), ensuring consistency of trajectories and error magnitudes. To prevent short-term spikes or inflated requests from distorting allocation, an effective upper allocation bound is imposed for each domain in window : 8 Here, and represent the exponentially weighted moving average and standard deviation of past demand , and determines the tightness of the cap. This data-driven bound provides stability by adapting to historical usage patterns while preserving the upper limit dictated by policy. Step 1. Control trigger . A domain enters the reallocation control phase when its observed demand at time surpasses the effective upper bound . This threshold represents the operational limit of allocation for each domain, dynamically defined based on historical usage and policy constraints. Formally, the trigger condition is given by Eq. ( 9 ): 9 Step 2. Reclaimable surplus computation . For each domain that exceeds the effective upper bound, the reclaimable amount is defined as the surplus of the predicted allocation share over the effective cap. Here, denotes the predicted allocation ratio obtained from the static estimator in Sect.  3.3, and is given by Eq. ( 8 ). The reclaimable surplus is computed as 10 which is positive only when the predicted allocation exceeds the effective cap. Domains with positive surplus values contribute to the redistribution pool described in Step 3. Step 3. Eligibility for redistribution . Domains with a positive reclaimable surplus are identified as eligible candidates for resource redistribution. This is formalized through the set definition in Eq. ( 11 ): Here, denotes the subset of domains where the reclaimable surplus is positive—i.e., domains that have received less allocation than permitted under their effective upper bound f. These domains are prioritized in the reallocation process to improve alignment with policy targets. 11 Step 4. Reallocation execution . In the subsequent allocation window , reclaimed resources are proportionally redistributed across all eligible under-allocated domains. The allocation weights are determined by the magnitude of each domain’s shortfall, . This proportional redistribution rule ensures that the reallocation process is both fair and policy-consistent, gradually reducing deviation from the effective target profile. In practice, two parameters—γ (loss-weight mixing) and κ (cap tightness)—control the balance between policy stability and operational flexibility. For reproducibility, we suggest setting γ within [0.3, 0.8], chosen via short-term validation on historical data, and selecting κ within [1.0, 2.0] to maintain a cap-hit rate of 5–15% while keeping GPU utilization above 90%. These ranges were found to yield stable performance across different workload profiles and system scales. Figure 3 illustrates the overall reallocation process as a sequential decision workflow. Fig. 3. Open in a new tab Reallocation process flow. Simulation results To evaluate the operational effectiveness of the proposed dynamic controller, we simulate resource allocation over a one-week horizon using time-series demand curves for six domains. Each week is segmented into 28 windows of 4-hour intervals. All demand and allocation values are normalized by the per-window total capacity, which is scaled to 1.0. This renders evaluation metrics such as MAE and RMSE of allocation ratios invariant to changes in absolute GPU capacity. For empirical calibration, the normalized curves are anchored to Neuron’s historical weekly GPU-hour delivery (aggregated and anonymized), ensuring consistency in trajectory and error magnitudes. Figure 4 presents the allocation ratio differences before and after applying the dynamic controller. In the uncontrolled baseline, certain domains exhibit large deviations from their target allocation ratios—for example, Materials and Chemistry are persistently over-allocated by − 17.6% points and − 6.1% points, respectively. After applying the control mechanism defined by Eq. ( 8 ) through (11), these deviations are substantially reduced. Most domain-level errors fall within ± 3% points. Engineering, for instance, improves from − 9.8%p to − 0.4%p, and Earth Science from − 9.0%p to − 0.8%p, indicating that the combined effect of effective capping and proportional reallocation significantly enhances policy alignment. Fig. 4. Open in a new tab Allocation ratio difference before and after control by domain (%). For additional benchmarking, we also implemented a demand-driven baseline that mirrors a fair-share or DRF-type scheduler, allocating GPU capacity purely in proportion to observed demand without policy weighting. This baseline produced MAE and RMSE values of 8.12 and 9.63%, respectively—comparable to the uncontrolled static case (8.03 and 9.59%)—whereas the proposed controller achieved much lower errors (1.30 and 1.66%) while maintaining GPU utilization above 92%. These results confirm that the proposed framework outperforms conventional demand-based scheduling in both policy alignment and operational stability. All error metrics in this section are reported out of sample under the rolling protocol; that is, ( ) are tuned on and evaluated at using the exogenous weights in Eq. ( 7 ). Quantitative evaluation corroborates these gains. Aggregated over time windows, the MAE decreased from 8.03 to 1.30%, and the RMSE decreased from 9.59% to 1.66%. The reduction reflects not only lower absolute error but also a more uniform improvement across domains, avoiding selective or biased correction. To assess performance neutrality, we additionally measure GPU utilization, average queue waiting time, and throughput under identical demand curves and capacity budgets. Mean GPU utilization remained above 92% after control; median queueing delay did not increase and modestly decreased in high-load intervals; and throughput was maintained. Expanded results are summarized in Table 5 , which reports MAE/RMSE together with the three operational metrics and 95% confidence intervals. Table 5. Quantitative evaluation of allocation error (MAE and RMSE metrics). Allocation ratio difference Allocation ratio difference(controlled) MAE 8.03% 1.30% RMSE 9.59% 1.66% Open in a new tab All error metrics are evaluated under the rolling protocol, where parameters ( ) are tuned using historical data from time steps and tested at step , using exogenous weights from Eq. ( 7 ). The overall quantitative gains are summarized in Table 5 . The MAE is reduced from 8.03% to 1.30%, and the RMSE from 9.59 to 1.66%. These reductions reflect not only lower average error but also more uniform improvements across domains, thereby avoiding selective or biased correction. Figures 5 and 6 visualize the domain-specific request and allocation shares over the 28 simulation windows. Figure 5 displays the raw request shares normalized by total capacity per window. Despite prior static optimization, observed request shares often deviate from policy targets. Several domains persistently over- or under-request relative to their assigned policy shares. Fig. 5. Open in a new tab Domain-specific resource demand curves (uncontrolled, T1–T28). Fig. 6. Open in a new tab Domain-specific resource demand curves (controlled and reallocated, T1–T28). Figure 6 illustrates the allocated shares after applying the control procedure. The controller successfully enforces the cap-and-redistribution rules: when demand exceeds the target, allocation tracks the cap; when demand falls below the target, allocated shares adjust downward accordingly without inducing idle capacity. This shows the controller’s flexibility in handling varying demand scenarios while preserving fairness. Figure 7 quantifies the volume of reallocated GPU-hours per domain, normalized to the weekly capacity. The results show that reclaimed capacity is not distributed uniformly across all domains but is systematically redirected to persistently under-served domains—notably Earth Science and Fusion. This demonstrates the effectiveness of the reclaim–reallocate mechanism in correcting long-term allocation imbalances. Fig. 7. Open in a new tab Quantity of reallocated resources per domain. To validate this interpretation, we performed an ablation comparison among Static-only, Dynamic-only, and Combined configurations. Only the Combined approach consistently maintains the policy-alignment potential at levels below the baseline across nearly all time windows, while minimizing cross-domain dispersion. Sensitivity analysis further confirms robustness: varying the loss-weight mixing parameter and cap tightness parameter leaves the qualitative outcomes unchanged. A quantitative sensitivity analysis over (γ, κ) is reported in Supplementary Information A.9, including policy-alignment and operational metrics. Convergence and strategic behavior To assess convergence toward policy alignment, we define the policy-alignment potential as the squared deviation between predicted allocations and target shares (Eq. ( 12 )): 12 As shown in Fig. 8 , where lower Vₜ indicates stronger policy alignment, the controlled trajectory remains consistently below the baseline and approaches zero in multiple windows. This confirms that the controller consistently drives allocation profiles toward the policy target vector , even under fluctuating workloads. Fig. 8. Open in a new tab Policy-alignment potential over time. As shown in Fig. 8 , where lower Vₜ indicates stronger policy alignment, the controlled trajectory remains consistently below the baseline and approaches zero in multiple windows. Lower values of indicate stronger policy alignment between predicted allocations and the policy target vector T. The controlled trajectory consistently achieves lower values compared to the baseline, confirming improved policy adherence. To examine robustness under potential manipulation, we simulate two stress scenarios for a single domain: (1) a short-lived demand spike of + 50% over 2 windows, and (2) a plateau of + 20% over 8 windows. Table 6 reports the resulting MAE and RMSE before and after applying the controller. In both scenarios, the controller substantially reduces deviation—by 40–45%—demonstrating its resilience against strategic inflation. Notably, the effective cap absorbs sudden spikes, while the priority-based redistribution scheme ensures fairness across domains. Table 6. Stress-test results under strategic demand inflation. Scenario MAE (baseline) RMSE (baseline) MAE (controlled) RMSE (controlled) Spike ×1.5 (2 windows) 0.0889 0.1300 0.0497 0.0767 Plateau + 20% (8 windows) 0.0896 0.1313 0.0510 0.0786 Open in a new tab Finally, to ensure operational neutrality, we examine system-level metrics such as GPU utilization, median queue waiting time, and throughput. Figure 9 illustrates the time-series trajectory of system-level GPU utilization across the 28 four-hour windows for both the static-only baseline (dashed line) and the proposed controlled configuration (solid line). The controlled trajectory remains consistently above 92% utilization, nearly overlapping with the baseline, indicating that throughput and queue delays are maintained without observable degradation. These results confirm that the proposed dynamic allocation mechanism enforces policy alignment while preserving core operational efficiency. Fig. 9. Open in a new tab System-level utilization vs. time. Discussion This study presents and empirically evaluates a policy-integrated resource allocation framework that internalizes the policy target vector within both the offline estimator and the runtime controller. Domain-specific characteristics are represented through two complementary signals: a structural-fit signal , computed as a cosine similarity against a system-level reference profile, and a demand-oriented signal . These signals are combined using interpretable weights and , estimated under a rolling out-of-sample protocol. To mitigate endogeneity concerns, only lagged information is used in the loss construction. From an operational perspective, the framework demonstrates improved policy alignment relative to the uncontrolled baseline. When combined with a capping mechanism and reclaim–reallocate rules, allocation deviations decrease under the evaluated demand traces. Under normalized-capacity conditions, MAE and RMSE are reduced while GPU utilization, throughput, and median queueing delay remain comparable to baseline levels. Figures 5 and 6 illustrate the pre- and post-intervention allocation trajectories, and Fig. 7 presents domain-level redistribution patterns. The policy-alignment potential remains below the baseline in most evaluation windows. Additional ablation and stress tests further indicate stable behavior across the tested scenarios. Practically, the framework provides a tunable interface between policy objectives and operational execution. The weights and regulate the balance between structural alignment and demand intensity. The blending parameter governs the incorporation of historical policy emphasis, while the cap tightness parameter controls enforcement flexibility. Because results are expressed in normalized allocation shares, they can be translated to absolute metrics such as weekly GPU-hours without affecting alignment measures. Although the simulation focuses on a single system with six domains, the estimator and controller operate on normalized shares, implying computational complexity that scales linearly with the number of domains. This structure allows potential extension to multi-cluster or heterogeneous environments. The framework may also be integrated with existing schedulers such as Slurm or PBS by incorporating the policy-aligned allocation vector as a weighting layer in queue priorities or QOS parameters. Such integration would not require modification of the underlying scheduling kernel. Several limitations remain. First, evaluation was conducted in a single-system simulation with a limited time horizon. Multi-cluster and longer-term deployments should be examined in future work. Second, the policy target vector was treated as static; in practice, it may evolve alongside national priorities and workload composition. Adaptive updating of and integration with demand forecasting models represent important extensions. Finally, while the controller mitigates over-allocation through effective capping, real-time systems may still face strategic behavior and noisy signals. These aspects warrant further investigation through live-system experimentation and adaptive control methods. Conclusion This study proposes a policy-integrated optimization framework for GPU resource allocation that combines a target-aware static estimator with a dynamic runtime controller. By incorporating the policy target vector and leveraging structural similarity and demand-intensity signals, the framework provides a structured mechanism for policy-aware reallocation under time-varying workloads. Across the evaluated demand scenarios, the approach reduces allocation deviation relative to both uncontrolled static optimization and a demand-driven baseline, while maintaining comparable levels of GPU utilization and queue-related metrics. Ablation and sensitivity analyses indicate that structural and demand-oriented signals play complementary roles, and the analytical formulation suggests stable behavior under the stated assumptions. The current implementation is limited to short evaluation horizons and simulated demand environments. Extensions to longer-term forecasting, multi-constraint optimization (e.g., energy or carbon considerations), and strategic interaction models represent promising directions for future research. Empirical validation in live operational settings will be important for assessing generalizability across heterogeneous infrastructure environments. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 (77.9KB, docx) Acknowledgements This study was funded by KISTI(K26L1M1C1). Author contributions All authors contributed to the study conception, design, material preparation, data collection and analysis. The first draft of the manuscript was written by Shim, Shim and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. Data availability The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Strohmaier, E., Dongarra, J., Simon, H. & Meuer, M. TOP500 Supercomputer Sites. https://www.top500.org (2024). 2. EuroHPC Joint Undertaking. Terms of Reference – Regular Access Call for Proposals. https://access.eurohpc-ju.europa.eu (2025). 3. National Energy Research Scientific Computing Center (NERSC). Allocation policies and procedures. https://www.nersc.gov/users/accounts/allocations/policies/ (2023). 4. RIKEN Center for Computational Science. Operation and Use of Supercomputer Fugaku. https://www.r-ccs.riken.jp/en/fugaku/ (2023). 5. Mo, Z., Xu, H. & Lau, W. C. Optimal Resource Efficiency with Fairness in Heterogeneous GPU Clusters. In Proceedings of the 25th International Middleware Conference (pp. 36–48). (2024). 6. Zeng, Q., Du, Y., Huang, K. & Leung, K. K. Energy-efficient resource management for federated edge learning with CPU-GPU heterogeneous computing. IEEE Trans. Wireless Commun. 20 (12), 7947–7962 (2021). [ Google Scholar ] 7. Ramamoorthi, V. AI-Driven Cloud Resource Optimization Framework for Real-Time Allocation. J. Adv. Comput. Syst. 1 (1), 8–15 (2021). [ Google Scholar ] 8. Tai, K. Y., Lin, F. Y. S. & Hsiao, C. H. An Integrated Optimization-Based Algorithm for Energy Efficiency and Resource Allocation in Heterogeneous Cloud Computing Centers. IEEE Access. 11 , 53418–53428 (2023). [ Google Scholar ] 9. Wang, S., Chen, S., Shi, Y., Yao, G. & Liu, M. AdaGap: An adaptive gap-aware resource allocation strategy for GPU sharing in heterogeneous clusters. Future Generation Comput. Syst. 173 , 107883 (2025). [ Google Scholar ] 10. Wang, X. et al. Kairos: Deterministic Scheduling Enhanced by User Collaboration for Deep Learning Workloads (IEEE Transactions on Parallel and Distributed Systems, 2025). 11. Wu, H., Wang, X., Luo, X., Shen, W. & Lin, W. MCG-Sched: Multi-cluster GPU scheduling for resource fragmentation reduction and load balancing. IEEE Trans. Parallel Distrib. Syst. 36 (12), 2789–2800 (2025). [ Google Scholar ] 12. Huang, J. et al. On Efficiency, Fairness and Security in AI Accelerator Resource Sharing: A Survey. ACM Comput. Surveys . 57 (9), 1–35 (2025). [ Google Scholar ] 13. Hart, D. L., Deems, S. L. & Herriott, L. T. From Vision to Evaluation: A Metrics Framework for the ACCESS Allocations Service. SN Comput. Sci. 5 (5), 490 (2024). [ Google Scholar ] 14. Johnson, C. et al. PRACE DECI (distributed European computing initiative) minisymposium. In Parallel Computing: Accelerating Computational Science and Engineering (CSE) (pp. 805–812). (IOS Press, 2014). Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1 (77.9KB, docx) Data Availability Statement The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (2.6 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top