IEEE Transactions on Cloud Computing最新文献

{"title":"ASP: An Adaptive Scaling Predictor With MoE-Enhanced Transformer for Cloud Resource Forecasting","authors":"Ye Liu;Jiongjiong Gu;Xianzhao Guo;Changping Ji;Yuzhao Wang;Weipeng Cao;Laizhong Cui;Zhong Ming","doi":"10.1109/TCC.2026.3713395","DOIUrl":"https://doi.org/10.1109/TCC.2026.3713395","url":null,"abstract":"Accurate cluster-level resource forecasting is pivotal for cloud capacity planning. Existing static heuristics fail to adapt to dynamic, heterogeneous workloads, causing resource waste or service disruptions. This paper proposes ASP (Adaptive Scaling Predictor), a learning-based framework optimized for cloud scenarios via three core innovations: a dual-path input fusing long-term periodic and short-term real-time features to capture global and local temporal patterns comprehensively; a MoE-Transformer architecture that enhances learning capacity and cross-scenario transferability, with sparse activation reducing deployment resource demands; a business-driven loss function balancing resource savings and task queuing rates, prioritizing service reliability alongside utilization.Experiments on Huawei Cloud and Alibaba Cloud production traces show ASP outperforms expert heuristics and state-of-the-art models, achieving 40.23% resource savings in Huawei Cloud dataset, a 0.01% task queuing rate, and a 10% underestimation ratio (vs. 20% for iTransformer) in Alibaba Cloud dataset.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1436-1447"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148902787","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":"FancyPIR: Stateless Batch Private Information Retrieval With Communication-Computation Trade-Offs","authors":"Pengfei Lu;Zengpeng Li;Mei Wang;Bingsheng Zhang","doi":"10.1109/TCC.2026.3706422","DOIUrl":"https://doi.org/10.1109/TCC.2026.3706422","url":null,"abstract":"Existing Private Information Retrieval (PIR) protocols perform reasonably well in practice. However, the communication overhead remains a significant obstacle for large-scale applications. A natural approach is to amortize the communication overhead by querying multiple entries simultaneously. Recent research has explored Batch PIR (Multi-query PIR) protocols that optimize asymptotic complexity in single-server scenarios. This paper presents <inline-formula><tex-math>$text {FancyPIR}$</tex-math></inline-formula>, a stateless batch PIR solution that attempts to trade off communication against computation without depending on the preprocessing hint. <inline-formula><tex-math>$text {FancyPIR}$</tex-math></inline-formula> aims to reduce the communication overhead while managing the computational cost. To this end, <inline-formula><tex-math>$text {FancyPIR}$</tex-math></inline-formula> employs two major strategies: first, <inline-formula><tex-math>$text {FancyPIR}$</tex-math></inline-formula> reduces the number of rotations in the initialization (offline) phase and the number of multiplication operations in the server response (online) phase, which decreases the computational burden; second, <inline-formula><tex-math>$text {FancyPIR}$</tex-math></inline-formula> packs multiple queries and merges responses using the vectorized homomorphic encryption during the online phase, significantly reducing the communication overhead. Specifically, for retrieving a batch of 256 entries from a database containing <inline-formula><tex-math>$2^{20}$</tex-math></inline-formula> entries (each 256 Bytes in size), <inline-formula><tex-math>$text {FancyPIR}$</tex-math></inline-formula> achieves a communication overhead reduction of 8.4<inline-formula><tex-math>$sim 104times$</tex-math></inline-formula> compared to state-of-the-art solutions (presented in IEEE S&P ’18 et al.), along with an improvement in initialization setup time by 1.9<inline-formula><tex-math>$sim 3times$</tex-math></inline-formula>. Considering communication and computation, <inline-formula><tex-math>$text {FancyPIR}$</tex-math></inline-formula> reduces server monetary costs by 8.5<inline-formula><tex-math>$sim 59times$</tex-math></inline-formula> for different database sizes.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1819-1831"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148902888","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":"SURE: A High-Performance, Efficient, and Secure Serverless Framework Based on Unikernels","authors":"Federico Parola;Shixiong Qi;Anvaya B. Narappa;K. K. Ramakrishnan;Fulvio Risso","doi":"10.1109/TCC.2026.3697819","DOIUrl":"https://doi.org/10.1109/TCC.2026.3697819","url":null,"abstract":"Current serverless platforms face substantial overhead from kernel-based networking and per-function sidecars. In addition, container-based runtimes suffer from excessive startup times and limited isolation. These limitations motivate the need for a more efficient and secure design. We present SURE, a unikernel-based serverless framework that combines rapid function startup with a high-performance, secure data plane. SURE enables distributed zero-copy communication through seamless integration of a userspace zero-copy TCP/IP stack (Z-stack) with intra-node shared memory processing. To eliminate the inefficiency of per-function sidecars, SURE introduces a lightweight library-based sidecar, which reduces CPU overhead by over two orders of magnitude compared to traditional userspace sidecars. For security, SURE leverages Intel’s Memory Protection Keys (MPK) to enforce fine-grained, page-level isolation in the shared memory data plane and to isolate the Trusted Computing Base (TCB) components in the function runtime (e.g., library-based sidecar, scheduler, etc) from untrusted user code. This is complemented by memory-pool-based security domains that isolate each function chain, ensuring scalability to large deployments. SURE further integrates a combination of binary inspection, W <inline-formula><tex-math>$oplus$</tex-math></inline-formula> X enforcement, and TCB-page blacklisting to prevent MPK privilege escalation within the single-address-space unikernel. These combined efforts create a more secure and efficient data plane with improved performance. Our evaluation shows that SURE improves throughput by 6×-8× compared to SPRIGHT, a high-performance serverless platform.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"2136-2153"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148902981","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":"Online Algorithms for Data Sharing-Aware Task Allocation in Edge Computing Systems","authors":"Sanaz Rabinia;Daniel Grosu","doi":"10.1109/TCC.2026.3697282","DOIUrl":"https://doi.org/10.1109/TCC.2026.3697282","url":null,"abstract":"Edge computing represents an approach in network architecture, allowing data processing closer to the data source or edge of the network. Transferring the data from the offloaded tasks by the end-user devices to edge devices can slow down the network due to increased latency and congestion. Many of these tasks offloaded to edge servers that perform computation to analyze the sensing data, may process the same data items. In order to reduce the traffic in the networks and the number of edge servers, we introduce task allocation algorithms that take into account the sharing of data among the tasks offloaded to the same server. Given that edge computing facilitates real-time processing, in this paper we design heuristic online algorithms for task allocation in edge computing systems. We perform an extensive performance analysis by comparing our proposed algorithms with several sharing-oblivious and sharing-aware baseline algorithms. The results show that our algorithms are able to reduce the amount of data transferred in the network by 30.2% to 92.8% and the number of utilized servers by 1% to 82.8% compared to the sharing-oblivious baseline algorithms. Compared to a sharing-aware baseline algorithm, our proposed algorithms are able to reduce the amount of data transferred in the network by 18.1% to 86.9% and the number of utilized servers by 35.4% to 88%. We also augment the proposed data sharing-aware algorithms with a local search phase that iteratively attempts to improve the solutions obtained by our data sharing-aware algorithms. It does this by exploring the neighborhood of the current solution and making small modifications. Our extensive experimental performance analysis shows that the algorithms augmented with local search reduce the number of utilized servers by 9.1% to 66.7% compared to the data sharing-aware online algorithms at the expense of a small increase in the amount of data transferred in the network and a small increase in execution time.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1989-2004"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148903103","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 Dynamic Microservice Scheduling Model With Elastic Scalability and Adaptive Awareness","authors":"Hongbo Wang;Bo Deng;Jinyu Zhang;Siqi Wang;Hong Zhu","doi":"10.1109/TCC.2026.3689123","DOIUrl":"https://doi.org/10.1109/TCC.2026.3689123","url":null,"abstract":"With the pervasive adoption of microservices (MS) architecture, resource scheduling has become a pivotal factor in ensuring system performance, stability, and elasticity. However, current MS resource scheduling approaches frequently overlook the complexities inherent in MS architecture, particularly the interdependencies between MS instances and the rapid change of dynamic workload patterns, thus fail to meet multiple but conflict performance objectives simultaneously. To address these challenges, this paper introduces a dynamic MS scheduling model DMSM-ESAAI which integrates elastic scalability and adaptive awareness with Kubernetes’ native scheduling and autoscaling pipeline at three levels: the instance and strategy levels of scheduling, and node level scaling. The experimental evaluations conducted on both real Kubernetes clusters and simulation environments demonstrated that the proposed model significantly outperforms traditional methods in terms of throughput, latency, and resource utilization, validating its effectiveness in dynamic MS environments.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1244-1261"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148903118","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":"Pikavolt: A Cloud-Native BlockDAG Consensus Protocol for Scalable Blockchain Networks","authors":"Hao Yin;Changling Zhou;Weiping Wen;Yiwei Liu","doi":"10.1109/TCC.2026.3707890","DOIUrl":"https://doi.org/10.1109/TCC.2026.3707890","url":null,"abstract":"Blockchain technology establishes a decentralized and trusted environment with a wide range of application scenarios. Cloud computing can serve as a foundational infrastructure to support the rapid deployment of scalable blockchain services. To overcome the inherent performance and capacity limitations of traditional blockchain structures, researchers have introduced the directed acyclic graph (DAG) to reconceptualize the data organization within blockchains. It gives rise to cloud-compatible BlockDAGs, broadening the application scope of blockchain technology and becoming a pioneering research direction in contemporary blockchain technology. In this paper, we propose a cloud-native consensus protocol called Pikavolt, designed for scalable blockchain networks built on the BlockDAG architecture. Pikavolt is capable of identifying convergent subgraphs in a divergent graph structure, thereby yielding stable consensus outcomes. The protocol first defines block height to stratify the DAG into layers and constructs block feature sets based on inter-node references to capture confirmation status. Subsequently, we develope an extended partitioning algorithm that adaptively segments blocks in each layer and merges them into a consensus-reaching subgraph. By leveraging cloud-native infrastructure, Pikavolt supports parallel execution and elastic deployment across distributed environments. Theoretical analysis demonstrates the algorithm’s effectiveness in identifying aggregated subgraphs, while experimental simulations show that Pikavolt outperforms the state-of-the-art schemes by about 2<inline-formula><tex-math>$sim$</tex-math></inline-formula>3 times in performance and achieves greater consistency.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"2045-2054"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148903121","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":"FedFEC: Tailored Learning via Multi-Level Knowledge Standardization and Fusion to Align Knowledge in Mobile Edge Computing","authors":"Yuxuan Long;Zhenyu Wang;Kai Xu;Rui Zhang","doi":"10.1109/TCC.2026.3691382","DOIUrl":"https://doi.org/10.1109/TCC.2026.3691382","url":null,"abstract":"As awareness of privacy protection for mobile devices increases, the application of federated learning in Mobile Edge Computing (MEC) has attracted widespread attention. Federated learning enables the cooperative training of a robust and capable global model across mobile devices without accessing private data. However, the communication overhead and model heterogeneity issues pose significant challenges for federated learning in MEC. Although federated distillation can serve as a viable approach to reduce communication overhead through knowledge distillation and facilitate knowledge transfer among heterogeneous models, the mismatch in knowledge amplitude due to differing parameter scales between local and global models can severely impact its effectiveness. To address these challenges, we propose a flexible knowledge standardization method to align the knowledge amplitude between teacher and student models. Additionally, we introduce a multi-level knowledge fusion federated distillation approach, FedFEC, designed to enhance the performance of federated distillation for MEC. FedFEC selects clients using the Wasserstein distance, aligns knowledge via a standardization function based on model size, and fuses multi-granularity knowledge for effective server-side distillation. Clients then update local personalized model based on multi-level global knowledge. Extensive experiments conducted on four datasets demonstrate that FedFEC outperforms all benchmarks across various scenarios of data heterogeneity and model architectures.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1846-1857"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148903161","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":"GEN-CSAF: A Generative Autonomy Framework for End-to-End Task Scheduling in Container-Based Systems","authors":"Lilu Zhu;Kai Huang;Yang Wang;Yingze Su","doi":"10.1109/TCC.2026.3707138","DOIUrl":"https://doi.org/10.1109/TCC.2026.3707138","url":null,"abstract":"Task scheduling plays a key role in the development of elastic, reliable and cost-efficient cloud computing paradigms. However, most prevailing approaches—typically relying on static configurations or manually crafted templates—suffer from dual limitations in parsing scheduling intents and adapting to cluster dynamics. Herein, we propose a generative scheduling autonomy framework (GEN-CSAF) to improve the scheduling performance of container-based cloud systems. This framework employs a novel decomposition of scheduling policies, which integrates pretrained code large language models (Code-LLMs) with a dual reflection mechanism to automatically generate interface-level policy codes. Initially, we exploit time-series heatmaps to represent global states and workload discrepancies within the current active cluster. Based on these representations, a structured API engine and a meta-catalog are developed to standardize interface encapsulation and consistent invocation, thereby transforming decision-making processes into semantically equivalent scheduling policy programs written in general-purpose languages. Subsequently, we design a retrieval-augmented generation (RAG) pipeline to implement semantic-aware reasoning of scheduling action codes. Additionally, we also adopt systematic reflection on actions through rule-based evaluation and runtime behavior-based evaluation to elide generic hallucinations from Code-LLMs. Experimental results show that, compared with the state-of-the-art process-oriented approaches, GEN-CSAF achieves a 15.37% increase in service-level objective compliance and a 21.95% increase in global resource utilization efficiency within a general architecture. In the end, we further validate our framework on a testbed comprising hybrid domestic architecture. For highly stochastic cloud-native workloads, GEN-CSAF demonstrates superior performance in cluster adaptability and system robustness.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1870-1887"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148903164","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":"Code-Aware and Load-Aware Resource Configuration for Serverless Functions via Deep Reinforcement Learning","authors":"Donghong Xu;Weijue Bu;Zhouliang Ye;Shilong Wu","doi":"10.1109/TCC.2026.3714912","DOIUrl":"https://doi.org/10.1109/TCC.2026.3714912","url":null,"abstract":"Serverless resource control must react to workload pressure without ignoring what the function actually computes. Existing profiling and telemetry-driven optimizers mainly observe runtime symptoms, which makes them slow to transfer across functions whose CPU, memory, and I/O behavior differ. CALO is a code-aware and load-aware controller that caches offline code embeddings, combines them with live telemetry in a 79-dimensional state, and selects memory, architecture, and timeout from a 48-action provider catalog. CALO is trained with a measurement-grounded batch-window simulator built from OpenWhisk warm, cold-start, burst, and idle-gap profiles, which supports simulator-based policy comparison while keeping the execution model tied to measured behavior. Across three seeds, five benchmarks, and four workload families under dual x64/ARM64 calibration, CALO improves mean raw reward from 0.513 to 0.569 over online Bayesian optimization and raises the 10% Conditional Value at Risk (CVaR<inline-formula><tex-math>$_{10}$</tex-math></inline-formula>) from 0.163 to 0.351. A matched Load-Only CALO ablation indicates a modest, workload-dependent code contribution: the largest mean gains occur on thumbnailer and image-recognition, while effects are mixed or small when runtime telemetry already explains the workload well. OpenWhisk validation covers 120 x64 fixed-configuration targets, yields 112 deployable configurations with 13.9% median warm-path error, and gives positive transfer evidence across eight focused adaptive-stress replay cases. Together, the simulator and OpenWhisk evidence indicate that code semantics can provide a complementary control signal within the tested settings, while online action selection remains lightweight at 0.124 ms median and 0.132 ms p95 on the local CPU used in this study.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1616-1629"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148902630","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":"Event-Driven Double Dueling Deep Q-Learning for Efficient Serverless Edge Function Offloading and Scheduling","authors":"Manish Pandey;Young-Woo Kwon","doi":"10.1109/TCC.2026.3703387","DOIUrl":"https://doi.org/10.1109/TCC.2026.3703387","url":null,"abstract":"Serverless edge computing combines the scalability and cost-efficiency of serverless platforms with the low-latency benefits of edge systems. However, the limited computational capacity of individual edge nodes and frequent cold starts often lead to long waiting times and missed deadlines. Although horizontal offloading can alleviate local overload, conventional heuristics rely on static metrics and neglect critical conditions such as queue buildup, warm containers, and memory allocation, which frequently shift bottlenecks to neighboring nodes. This paper proposes a cooperative serverless edge architecture that integrates intelligent horizontal offloading with an adaptive in-node scheduling layer. A Double Dueling Deep Q-Learning (D3QL) agent guides task placement based on real-time feedback, while the scheduler at the chosen node prioritizes requests to increase warm-start reuse and protect deadline-sensitive workloads. The scheduler’s decisions flow back into the agent, forming a continuous improvement loop between global and local control. Evaluations using production Azure and Alibaba traces show that our method consistently outperforms baseline approaches, achieving higher task completion and guarantee rates, reduced memory consumption, and improved warm-start ratios. Using a one-hour Azure invocation trace on OpenWhisk, our method achieves 67.46% higher task completion than local execution.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1802-1818"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148902679","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}