{"title":"Architecture of Japan’s NII research data cloud as a national open science infrastructure","authors":"Ikki Fujiwara","doi":"10.1016/j.future.2026.108933","DOIUrl":"https://doi.org/10.1016/j.future.2026.108933","url":null,"abstract":"NII Research Data Cloud (NII RDC) is Japan’s national infrastructure that operationalizes open science by integrating research data management, publication, and discovery. It comprises three lifecycle-aligned services: GakuNin RDM for secure, project-level collaboration and integrity-oriented workflows; JAIRO Cloud for institutional repository-based dissemination of research outputs; and CiNii Research for cross-domain discovery via a knowledge-graph approach. Developed by the National Institute of Informatics (NII) and its Research Center for Open Science and Data Platform (RCOS) from 2017 and operated at scale from 2021, NII RDC is embedded in national coordination efforts and supports universities in end-to-end handling of research outputs. This paper overviews the platform and maps GakuNin RDM to Research Data Implementation Architecture (RDIA) building blocks.","PeriodicalId":55132,"journal":{"name":"Future Generation Computer Systems-The International Journal of Escience","volume":"17 1","pages":""},"PeriodicalIF":7.5,"publicationDate":"2026-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149423831","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"CPHA-FB: Cascade-aware Proactive Holistic Autoscaling with dual-layer closed-loop feedback for large-scale microservice systems","authors":"Tahira Amin, Muhammad Asif, Wang-Cheol Song","doi":"10.1016/j.future.2026.108926","DOIUrl":"https://doi.org/10.1016/j.future.2026.108926","url":null,"abstract":"Cascade failures in microservice-based cloud systems remain a major operational challenge, where overload in one service propagates through dependency chains, triggering widespread SLA violations before reactive autoscalers can respond. Existing proactive approaches either model services independently, ignoring inter-service propagation, or lack mechanisms to correct scaling decisions after deployment. This paper presents CPHA-FB, a closed-loop autoscaling framework integrating spatio-temporal forecasting, graph-based cascade propagation, and adaptive runtime feedback. A Spatio-Temporal Graph Neural Network (ST-GNN) over dynamically weighted dependency graphs jointly forecasts demand across all services, reducing prediction error by 61.9% relative to a graph-unaware LSTM baseline, and predicted overload risk is propagated through the graph to provision downstream services before cascade escalation. Two feedback mechanisms improve runtime robustness without retraining: a confidence-aware safety margin that adjusts scaling aggressiveness according to forecasting reliability, and a per-service boost policy that adds capacity upon SLA degradation and retracts excess replicas after recovery. Evaluated on the Alibaba Cluster Trace 2021 dataset (987 million records, 150 microservices), CPHA-FB reduces SLA violations by 27.3% under normal load and by 63.6% relative to CPHA under cascade stress, reaching 0.0184% violations under full cascade injection, an 88.2% reduction relative to HPA, at 8.4% additional resource overhead. Reported results use trace-driven simulation, with dependency weights computed over a forward-shifted correlation window; a past-only variant is evaluated in the paper.","PeriodicalId":55132,"journal":{"name":"Future Generation Computer Systems-The International Journal of Escience","volume":"22 1","pages":""},"PeriodicalIF":7.5,"publicationDate":"2026-11-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149423880","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Sergio Infante, Julia Robles, Alejandro Megias, Luis Llopis, Fernando Díaz, Cristian Martín, Manuel Díaz
{"title":"An open-source agent-based digital twin architecture built on OpenTwins","authors":"Sergio Infante, Julia Robles, Alejandro Megias, Luis Llopis, Fernando Díaz, Cristian Martín, Manuel Díaz","doi":"10.1016/j.future.2026.108924","DOIUrl":"https://doi.org/10.1016/j.future.2026.108924","url":null,"abstract":"Digital twins (DT) have evolved into complex, software-intensive systems that require continuous synchronization, orchestration, and runtime adaptation. However, many existing DT platforms still rely on monolithic architectures that limit their scalability, dynamic integration, and plug-and-play deployment of heterogeneous decision-making logic. To overcome these limitations, this paper presents an open-source, agent-based DT architecture, built as an extension of the OpenTwins framework, which elevates Multi-Agent Systems (MAS) to the core orchestration layer. The contribution introduces an Agent Management Service and a set of reusable agent templates that abstract infrastructural complexities, such as database connections, asynchronous FMI co-simulations, and MQTT communication, through Kubernetes-orchestrated containerization. The capabilities of the architecture are rigorously validated through a dual approach. First, a highly distributed swarm intelligence scenario is deployed as a stress-test to demonstrate the system’s decentralized coordination, low-latency asynchronous communication, and emergent behavior. Second, the platform is instantiated in a real-world dynamic environment (a research building<ce:cross-ref ref><ce:sup loc=\"post\">2</ce:sup></ce:cross-ref><ce:footnote><ce:label>2</ce:label><ce:note-para view=\"all\">Details of the building will be revealed after review.</ce:note-para></ce:footnote>), which acts as a comprehensive workload. The results confirm the modularity, fault tolerance, and agent portability of the platform by demonstrating how drastically different computational paradigms, ranging from rule-based search algorithms to heavy Deep Q-Network (DQN) models, can be dynamically swapped and seamlessly orchestrated without modifying the underlying DT infrastructure.","PeriodicalId":55132,"journal":{"name":"Future Generation Computer Systems-The International Journal of Escience","volume":"34 1","pages":""},"PeriodicalIF":7.5,"publicationDate":"2026-11-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149424042","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Riccardo Cantini, Alessio Orsino, Domenico Talia, Paolo Trunfio
{"title":"A proxy-based interpretable model for energy estimation in Edge AI via in-context learning","authors":"Riccardo Cantini, Alessio Orsino, Domenico Talia, Paolo Trunfio","doi":"10.1016/j.future.2026.108921","DOIUrl":"https://doi.org/10.1016/j.future.2026.108921","url":null,"abstract":"Edge AI enables machine learning training and inference directly on resource-constrained devices, reducing latency, bandwidth usage, and privacy risks. However, the computational, memory, and communication demands of modern models make energy efficiency a key bottleneck for real-world deployment. Existing energy modeling frameworks often lack interpretability, limiting insight into <ce:italic>why</ce:italic> energy is consumed. This work introduces <ce:italic>PRIME-Edge</ce:italic> (<ce:italic>PRoxy-based Interpretable Model for Energy consumption in Edge AI</ce:italic>), an interpretable, proxy-based methodology for modeling energy consumption in Edge AI workloads. Energy usage is expressed as a weighted combination of learnable proxies capturing three components: <ce:italic>computation</ce:italic>, <ce:italic>data access</ce:italic>, and <ce:italic>communication</ce:italic>. Coupled with explainable AI techniques, this decomposition enables component-wise and feature-level interpretability, revealing dominant energy drivers across algorithms, datasets, and deployment configurations. Energy prediction is performed using tabular foundation models via in-context learning (ICL), achieving high accuracy and sample efficiency without task-specific retraining. The execution shots conditioning the ICL models are retrieved from an informative execution dataset constructed via design of experiments (DoE)–driven sampling, enabling robust, space-filling exploration while reducing measurement costs. PRIME-Edge is validated through extensive experiments on both classical and deep learning workloads in single- and multi-device settings on Raspberry Pi–based edge platforms. Results demonstrate the effectiveness of DoE-driven sampling, high sample efficiency of ICL models, and accurate energy estimation across diverse workloads. Overall, PRIME-Edge provides accurate energy estimation along with interpretable and actionable insights, enabling energy-aware practices for sustainable edge intelligence.","PeriodicalId":55132,"journal":{"name":"Future Generation Computer Systems-The International Journal of Escience","volume":"23 1","pages":""},"PeriodicalIF":7.5,"publicationDate":"2026-11-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149424097","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Othmane Belarbi, Theodoros Spyridopoulos, Yogha Restu Pramadi, Pietro Carnelli, Aftab Khan
{"title":"A comparative analysis of hyperdimensional computing and deep learning for network intrusion detection in IoT environments","authors":"Othmane Belarbi, Theodoros Spyridopoulos, Yogha Restu Pramadi, Pietro Carnelli, Aftab Khan","doi":"10.1016/j.future.2026.108920","DOIUrl":"https://doi.org/10.1016/j.future.2026.108920","url":null,"abstract":"The rapid proliferation of Internet of Things (IoT) devices has enabled transformative applications across smart homes, healthcare, and industrial automation. However, the limited computational resources of these devices and their deployment in large, distributed networks expose them to significant cybersecurity risks. Although Deep Learning (DL) has significantly improved intrusion detection rates, its high computational complexity and significant communication overhead render it unsuitable for many resource-constrained IoT devices, particularly within Federated Learning (FL) paradigms. To address this challenge, we propose FedHDC-IDS, the first application-driven framework that integrates Hyperdimensional Computing (HDC) with Radial Basis Function (RBF) encoding into the FL paradigm. FedHDC-IDS replaces complex gradient-based optimisations with lightweight hypervector operations to significantly reduce training complexity, memory footprints, and communication costs. We evaluate the framework against state-of-the-art models across three diverse IoT/IIoT datasets, under both centralised and federated settings, for binary and multi-class classification. The experimental results demonstrate that FedHDC-IDS achieves up to a 74% reduction in model size and up to a 32% reduction in training time compared with DL models, while maintaining comparable detection accuracy. These findings highlight the potential of HDC as an efficient alternative to DL for intrusion detection in IoT environments.","PeriodicalId":55132,"journal":{"name":"Future Generation Computer Systems-The International Journal of Escience","volume":"14 1","pages":""},"PeriodicalIF":7.5,"publicationDate":"2026-11-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149424043","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Sibo Qiao, Huimin Guo, Min Wang, Cong Liu, Mohammad Shabaz, Hamad Aldawsari, Xiao He
{"title":"An entropy-value deep reinforcement learning framework for dynamic task scheduling in distributed IoT environments","authors":"Sibo Qiao, Huimin Guo, Min Wang, Cong Liu, Mohammad Shabaz, Hamad Aldawsari, Xiao He","doi":"10.1016/j.future.2026.108918","DOIUrl":"https://doi.org/10.1016/j.future.2026.108918","url":null,"abstract":"In Internet of Things environments, stochastic task arrivals and heterogeneous virtual machine resources introduce nonstationarity and uncertainty. The resulting state drift and information latency bias the estimated values, leading to unstable convergence and policy degradation in large-scale schedulers. To address these challenges, we propose an Entropy-Value Cooperative Deep Reinforcement Learning framework (EV-D3QN), which integrates two key mechanisms: Entropy-Driven Exploration Reshaping (ERER) and Temporal-Consistency Memory Optimization (TCMO). The ERER adaptively regulates exploration intensity according to policy entropy and introduces a temperature-based acceptance probability to balance exploration and exploitation, addressing explore-exploit imbalance in nonstationary environments, thus improving decision reliability and stability. The TCMO dynamically reweights experiences based on the temporal evolution of temporal-difference error and the sample life cycle, addressing priority erosion and sample staleness caused by information latency and distribution shift, hence improving sample efficiency and enhancing convergence robustness. Extensive simulations demonstrate that the proposed EV-D3QN consistently outperforms the selected baseline algorithms under the evaluated experimental conditions. Compared with the best-performing baseline among those evaluated, EV-D3QN reduces the average makespan, completion-time variance, and total energy consumption by 4.0%, 7.2%, and 5.7%, respectively, thereby achieving high scheduling efficiency and strong robustness in nonstationary IoT systems.","PeriodicalId":55132,"journal":{"name":"Future Generation Computer Systems-The International Journal of Escience","volume":"30 1","pages":""},"PeriodicalIF":7.5,"publicationDate":"2026-11-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149424098","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Jiahui Wang, Zheyi Chen, Hailin Feng, Xing Liu, Kun Hua, Jia Lu
{"title":"FedSHAPE: Sharpness-guided trajectory harmonization and parameter equalization for federated learning under data scarcity and domain skew","authors":"Jiahui Wang, Zheyi Chen, Hailin Feng, Xing Liu, Kun Hua, Jia Lu","doi":"10.1016/j.future.2026.108927","DOIUrl":"https://doi.org/10.1016/j.future.2026.108927","url":null,"abstract":"Federated learning (FL) has emerged as a promising paradigm for privacy-preserving collaborative training. However, under the intertwined challenges of severe data scarcity (e.g., 1% sampling rate) and non-IID domain skew, existing FL approaches often experience substantial performance degradation. To address this issue, we introduce FedSHAPE (Federated Sharpness-aware Harmonized Aggregation and Parameter Equalization), a framework combining Stable Parameter Optimization (SPO) and Harmonized Aggregation and Equalization (HAE). By investigating the performance bottlenecks in such restrictive settings, we identify two coupled factors: standard local optimizers rapidly converge into domain-specific sharp minima (severe overfitting), rendering subsequent global feature alignment biased and suboptimal. Accordingly, FedSHAPE formulates a spatial–temporal optimization pipeline: it first executes spatial geometry regularization at the client level to suppress local overfitting noise, providing the design motivation for subsequent temporal trajectory harmonization and diversity-aware global feature alignment. Extensive experiments demonstrate that FedSHAPE consistently improves both generalization robustness and cross-domain fairness under highly restricted data conditions. Code is available at: <ce:inter-ref xlink:href=\"https://github.com/JiahuiWang0/FedSHAPE\" xlink:type=\"simple\">https://github.com/JiahuiWang0/FedSHAPE</ce:inter-ref>.","PeriodicalId":55132,"journal":{"name":"Future Generation Computer Systems-The International Journal of Escience","volume":"12 1","pages":""},"PeriodicalIF":7.5,"publicationDate":"2026-11-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149424093","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"An architecture-aware heterogeneous multigrid velocity solver for CitcomCU on the new-generation Tianhe supercomputer","authors":"Xiaoxiong Zhu, Yian Cui, Qinglin Wang, Jie Liu","doi":"10.1016/j.future.2026.108928","DOIUrl":"https://doi.org/10.1016/j.future.2026.108928","url":null,"abstract":"Large-scale mantle-convection simulations repeatedly solve memory-bound sparse velocity–pressure systems, making multigrid velocity solves a major performance bottleneck. We present an MT-3000-oriented heterogeneous multigrid solver for CitcomCU on the new-generation Tianhe supercomputer. Without changing the governing equations, finite-element discretization, Uzawa iteration, or multigrid hierarchy, the design reorganizes block Gauss–Seidel smoothing and sparse matrix–vector multiplication using regular nodewise storage, hierarchical data streaming, and overlapped host–accelerator execution. The evaluation covers controlled single-node comparisons for two physical cases and three global grids, application-level timing, kernel bandwidth analysis, numerical-consistency tests, and large-scale strong and weak scaling. Validation runs for isoviscous Case 1a and temperature-dependent Newtonian Case 2 completed under the prescribed convergence criteria, and the computed fields remain consistent with the unmodified CitcomCU 1.0.3 reference results. Relative to a fully threaded 16-core CPU baseline, the complete heterogeneous implementation achieves speedups of up to <mml:math altimg=\"si1.svg\" display=\"inline\"><mml:mrow><mml:mn>2.22</mml:mn><mml:mo>×</mml:mo></mml:mrow></mml:math> in mean aggregate multigrid (MG) time per simulation step and <mml:math altimg=\"si2.svg\" display=\"inline\"><mml:mrow><mml:mn>2.21</mml:mn><mml:mo>×</mml:mo></mml:mrow></mml:math> in complete Uzawa time. Across both cases and all three grids, the complete application achieves measured speedups of 1.21–<mml:math altimg=\"si3.svg\" display=\"inline\"><mml:mrow><mml:mn>1.63</mml:mn><mml:mo>×</mml:mo></mml:mrow></mml:math> relative to the same CPU baseline. Weak scaling retains 50.98% efficiency at 12,800 processes; the remaining efficiency loss is consistent with increasing communication and synchronization overhead.","PeriodicalId":55132,"journal":{"name":"Future Generation Computer Systems-The International Journal of Escience","volume":"14 1","pages":""},"PeriodicalIF":7.5,"publicationDate":"2026-11-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149424092","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Francisco Carter, Nancy Hitschfeld, Cristóbal A. Navarro
{"title":"Mìmir: An open-source library for real-time interactive visualization of CUDA programs","authors":"Francisco Carter, Nancy Hitschfeld, Cristóbal A. Navarro","doi":"10.1016/j.future.2026.108922","DOIUrl":"https://doi.org/10.1016/j.future.2026.108922","url":null,"abstract":"Real-time visualization of scientific applications is a valuable feature for many engineering and scientific research fields, as it allows researchers to visually assess the quality and correctness of the computational models involved during the execution, as well as to generate visual samples for communicating results. For iterative GPU-based applications based on platforms such as CUDA, existing visualization approaches based on copying to RAM or disk storage are not feasible in many cases, as they involve large memory transfers between GPU and CPU at each time-step. This round-trip for data produces a significant penalty for both performance and interactivity. On the other hand, implementing in-situ real-time visualization tools for GPU-based codes is a challenging task, as it involves dealing with (i) low-level integration of graphics APIs (e.g., OpenGL and Vulkan) with the parallel GPU code, (ii) a careful and efficient handling of memory spaces and (iii) finding a balance between rendering and computing as both tasks use the GPU resources. In this work we present Mìmir, an open source C++ library allows visualizing CUDA-based applications with Vulkan using a synchronized, zero-copy memory mapping between CUDA and Vulkan memory. Internally, Mìmir encapsulates the advanced and often complex features of Vulkan and reduces it to a compact API that can be called from within CUDA programs, allowing a programmer to describe visualizations in terms of CUDA/C++ data structures. This abstraction scheme allows generating visualizations with minimal alteration over the original application code, needing only to replace the GPU memory allocation lines of the data to be visualized by the API calls provided by Mìmir, among other optional changes for visual enhancements. Experimental benchmarks show that this library can generate real-time visualizations up to <mml:math altimg=\"si1.svg\" display=\"inline\"><mml:mrow><mml:mn>64.7</mml:mn><mml:mo>×</mml:mo></mml:mrow></mml:math> faster for local rendering of particle-based visualizations, and up to <mml:math altimg=\"si2.svg\" display=\"inline\"><mml:mrow><mml:mn>58.5</mml:mn><mml:mo>×</mml:mo></mml:mrow></mml:math> for memory-bound grid rendering, such as cellular automata. Mìmir also supports remote-rendering and level of detail (LOD) for large HPC workloads running on headless compute servers. Experiments using top-tier data center GPUs show that the library can enable real-time visualization of large scale simulations in the orders of billions of elements, using path-tracing methods to render up to 9, 7 million voxels representing 16 billion particles with <mml:math altimg=\"si3.svg\" display=\"inline\"><mml:mrow><mml:mi>L</mml:mi><mml:mi>O</mml:mi><mml:mi>D</mml:mi><mml:mn>256</mml:mn></mml:mrow></mml:math>.","PeriodicalId":55132,"journal":{"name":"Future Generation Computer Systems-The International Journal of Escience","volume":"4 1","pages":""},"PeriodicalIF":7.5,"publicationDate":"2026-11-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149424096","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"A unified measurement framework for scalability and resilience in cloud-native systems","authors":"Amro Al-Said Ahmad, Tobias Hoßfeld","doi":"10.1016/j.future.2026.108930","DOIUrl":"https://doi.org/10.1016/j.future.2026.108930","url":null,"abstract":"Scalability and resilience are interrelated quality attributes in cloud-native computing; however, they are commonly analysed separately despite this interdependence under changing workloads. Scalability mechanisms that are not considered adequate resilience measures may worsen failure propagation, leading to the scaling of faulty components or overwhelming dependent services. On the other hand, resilience mechanisms without effective scalability strategies can lead to resource exhaustion, degrade performance, or cause service unavailability as workloads grow. This work proposes a unified framework for scalability and resilience in cloud-native systems. It builds on scalability measurement frameworks that compute the scalability of cloud-native systems using empirical data and develops statistically grounded confidence intervals to quantify uncertainty in scalability evaluations. It then extends this foundation to introduce the resilience index, which captures systems under failure conditions, and integrates both qualities into a composite scalability-resilience index that provides a single evaluation metric focused on these interdependent attributes. The proposed framework is validated in a real cloud environment through more than 75 h of controlled scalability and resilience benchmark evaluations under diverse workload scenarios, including growth and burst states. Benchmarking results demonstrate that scalability and resilience are tightly coupled: systems that scale efficiently without resilience safeguards exhibit degraded reliability, while resilience mechanisms without scalable design fail under sustained load. Benchmarking evaluations further show that resilience degrades more sharply than scalability under load (resilience index <mml:math altimg=\"si1.svg\" display=\"inline\"><mml:mo>≈</mml:mo></mml:math> 0.21–0.43 versus scalability index <mml:math altimg=\"si1.svg\" display=\"inline\"><mml:mo>≈</mml:mo></mml:math> 0.51–0.53), and an instance-provisioning analysis identifies provisioning behaviour as the shared mechanism linking the two attributes. By providing a unified and statistically robust methodology, this work supports more informed design, benchmarking, and optimisation of cloud-native systems operating under evolving workload demands.","PeriodicalId":55132,"journal":{"name":"Future Generation Computer Systems-The International Journal of Escience","volume":"16 1","pages":""},"PeriodicalIF":7.5,"publicationDate":"2026-11-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149423835","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}