IEEE Transactions on Cloud Computing最新文献_第4页

Dynamic Distributed Encrypted Data Sharing With Lightweight Integrity Checking via Blockchain Systems 基于区块链系统的轻量级完整性检查的动态分布式加密数据共享

IF 5.3 2区计算机科学

IEEE Transactions on Cloud Computing Pub Date : 2026-07-01 Epub Date: 2026-06-29 DOI: 10.1109/TCC.2026.3708084

Bingyun Liu;Xiaojun Zhang;Yuan Zhang;Rang Zhou;Juncai Chen

{"title":"Dynamic Distributed Encrypted Data Sharing With Lightweight Integrity Checking via Blockchain Systems","authors":"Bingyun Liu;Xiaojun Zhang;Yuan Zhang;Rang Zhou;Juncai Chen","doi":"10.1109/TCC.2026.3708084","DOIUrl":"https://doi.org/10.1109/TCC.2026.3708084","url":null,"abstract":"With the explosive growth of various massive data, cloud computing relies on powerful storage services and computing resources to manage outsourced data. As most of data involve sensitive information, any tampering data will cause serious consequences, the integrity of outsourced data has become the most concerning security issue. Besides, it is of great importance to enable the sharing of outsourced data encrypted by different data users. To this end, we devise the dynamic distributed encrypted data sharing with lightweight integrity checking (IC-DEDS) mechanism via blockchain systems. We implement dynamic group division of members and secure member authentication through cryptographic accumulator and distributed authentication information. Further, we introduce double trap chamelen hash to implement the dynamic addition and deletion of group members. IC-DEDS designs joint public keys to achieve distributed encrypted data sharing among dynamic group members. IC-DEDS contributes to transparently integrity checking for outsourced encrypted data, and identity tracking of shared behaviors by utilizing blockchain and smart contracts, avoiding malicious behaviors and guaranting the traceability of shared data. Meanwhile, IC-DEDS ensures identity anonymity, revokes malicious users, achieves lightweight batch encrypted data integrity checking. The security analysis and performance evaluation demonstrate the feasibility of the mechanism in the secure deployment of cloud storage via blockchain systems.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1704-1719"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148902631","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}

引用次数: 0

A Two-Stage Framework for Cloud Service Combinatorial Optimization: Cognitive-Inspired Co-Evolutionary Dual-Population Search and Contrastive Compensation Ranking 云服务组合优化的两阶段框架：认知启发的协同进化双种群搜索和对比补偿排序

IF 5.3 2区计算机科学

IEEE Transactions on Cloud Computing Pub Date : 2026-07-01 Epub Date: 2026-07-03 DOI: 10.1109/TCC.2026.3710009

Hao Lu;Yang Song;Xingwei Wang;Yan Wang;Pengbo Li;Bo Yi;Min Huang;Keqin Li

{"title":"A Two-Stage Framework for Cloud Service Combinatorial Optimization: Cognitive-Inspired Co-Evolutionary Dual-Population Search and Contrastive Compensation Ranking","authors":"Hao Lu;Yang Song;Xingwei Wang;Yan Wang;Pengbo Li;Bo Yi;Min Huang;Keqin Li","doi":"10.1109/TCC.2026.3710009","DOIUrl":"https://doi.org/10.1109/TCC.2026.3710009","url":null,"abstract":"The rapid proliferation of heterogeneous cloud services has made Cloud Service Combinatorial Optimization (CSCO) a fundamental challenge for users and cloud service providers (CSPs) in large-scale computing ecosystems. Its complexity arises from the exponential growth of the decision space due to numerous service providers and tightly coupled resource constraints, making it difficult to efficiently identify high-quality CSP combinations. Moreover, intrinsic conflicts among cost, Quality of Service (QoS), and resource utilization further complicate the search for balanced solutions. In addition, existing multi-attribute decision-making approaches often fail to capture the compensatory trade-offs between price and QoS, resulting in limited interpretability and weak discrimination among candidate solutions. To address these challenges, we propose a two-stage cloud service combinatorial optimization framework that integrates evolutionary search and decision-oriented ranking. In Stage I, a cognitive-inspired co-evolutionary dual-population multi-objective optimization algorithm (CDMOA) is developed to efficiently explore the large CSCO search space. By coordinating a Strategist Population that emphasizes feasibility and cost-effectiveness with an Explorer Population that promotes solution diversity, CDMOA generates high-quality and well-distributed Pareto-optimal solutions. In Stage II, a bidirectional price–QoS contrastive compensation (BPQC) method is introduced to rank the obtained Pareto solutions. By integrating global–local contrast analysis with a bounded price–QoS compensation mechanism, BPQC operationalizes relative price–QoS trade-offs and provides interpretable rankings for decision support. Extensive experiments demonstrate that the proposed framework achieves competitive optimization performance and more effective decision ranking compared with several state-of-the-art baselines.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1358-1373"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148902641","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}

引用次数: 0

CDSSA: Cost-Effective Distributed Storage System Architecture for Ceph CDSSA：面向Ceph的高性价比分布式存储系统架构

IF 5.3 2区计算机科学

IEEE Transactions on Cloud Computing Pub Date : 2026-07-01 Epub Date: 2026-03-14 DOI: 10.1109/TCC.2026.3693306

Myoungwon Oh;Sungmin Lee;Yangwoo Roh;Jeong-Uk Kang;Silwan Chang

{"title":"CDSSA: Cost-Effective Distributed Storage System Architecture for Ceph","authors":"Myoungwon Oh;Sungmin Lee;Yangwoo Roh;Jeong-Uk Kang;Silwan Chang","doi":"10.1109/TCC.2026.3693306","DOIUrl":"https://doi.org/10.1109/TCC.2026.3693306","url":null,"abstract":"This paper presents a cost-effective distributed storage system architecture, named CDSSA, to handle a massive amount of data while serving good performance at a minimum cost. We attempted to build such a system on Ceph, a widely used open-source distributed storage system, employing deduplication with cost-attractive flash storage devices. However, two fundamental challenges are found to achieve the goal: (1) unnecessary additional writes and overwrites, caused by tightly coupled reference metadata with data chunk, make it difficult to exploit the affordable flash storage devices effectively, and (2) fragmentation due to distributed deduplication results in significant performance degradation and space overhead. To this end, CDSSA introduces two distinctive techniques: WORM architecture to avoid write amplification and enable affordable flash devices, and object packing to minimize fragmentation, improving both performance and space efficiency. We implement CDSSA on Ceph while maintaining its existing consistency and reliability. Our evaluation results show that CDSSA achieves a total storage space reduction of over 55% and improves read throughput by 75% on average.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1585-1601"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148902650","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}

引用次数: 0

TASADOR: A Machine Learning Framework for Network Bandwidth to CPU Translations 一个网络带宽到CPU转换的机器学习框架

IF 5.3 2区计算机科学

IEEE Transactions on Cloud Computing Pub Date : 2026-07-01 Epub Date: 2026-06-18 DOI: 10.1109/TCC.2026.3705455

Kyungwoon Lee;Sooyeon Woo;Jaehyun Hwang;Chuck Yoo

{"title":"TASADOR: A Machine Learning Framework for Network Bandwidth to CPU Translations","authors":"Kyungwoon Lee;Sooyeon Woo;Jaehyun Hwang;Chuck Yoo","doi":"10.1109/TCC.2026.3705455","DOIUrl":"https://doi.org/10.1109/TCC.2026.3705455","url":null,"abstract":"In virtualized cloud environments, meeting the network bandwidth requirements of virtual machines (VMs) is challenging due to the lack of accurate methods for translating target network bandwidth into precise CPU allocations. This translation is non-trivial, as outcomes vary depending on factors such as workloads, message sizes, and hardware configurations. For instance, our experiments reveal that for the same CPU allocation, network bandwidth can vary between <inline-formula><tex-math>$0.4times$</tex-math></inline-formula> and <inline-formula><tex-math>$3.5times$</tex-math></inline-formula>. In this paper, we introduce TASADOR, a machine learning (ML) framework for network bandwidth to CPU translations. TASADOR predicts “appropriate” CPU budgets for new workloads to meet their bandwidth requirements. We demonstrate that TASADOR completes data collection and training in <inline-formula><tex-math>${sim }16$</tex-math></inline-formula> minutes while achieving target bandwidth accurately with <inline-formula><tex-math>${sim }6.5times$</tex-math></inline-formula> less CPU utilization compared to existing schemes. Our experiments also confirm that TASADOR is effective across diverse workloads, message sizes, and hardware configurations.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"2154-2167"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148902845","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}

引用次数: 0

Cascading Segmented and Successive Inheritance Restoration Methods for NFV-Based Fog Computing 基于nfv的雾计算级联分割和连续继承恢复方法

IF 5.3 2区计算机科学

IEEE Transactions on Cloud Computing Pub Date : 2026-07-01 Epub Date: 2026-03-01 DOI: 10.1109/TCC.2026.3689769

Nazli Siasi;Mohammed Jasim

{"title":"Cascading Segmented and Successive Inheritance Restoration Methods for NFV-Based Fog Computing","authors":"Nazli Siasi;Mohammed Jasim","doi":"10.1109/TCC.2026.3689769","DOIUrl":"https://doi.org/10.1109/TCC.2026.3689769","url":null,"abstract":"Fog computing enhanced by network function virtualization (NFV) is pivotal for latency-sensitive applications such as autonomous vehicles and real-time healthcare. However, node failures attributed to hardware malfunctions, cyberattacks, or natural disasters disrupt services, degrade performance, and violate the service function chain (SFC) integrity. Existing solutions suffer from progressive redundancy, lack ultra-low latency restoration, efficient resource utilization, and scalability under high traffic. Hence, this work proposes novel post-fault restoration methods, termed as cascading segmented (CS) and successive inheritance (SI). The first partitions SFCs into localized segments and maps them onto fog nodes, while embedding their backup VNFs between the segments. The second employs a rolling mapping strategy for single-hop failover, where each node hosts both a primary VNF and the backup of its predecessor VNF. If failure occurs, restoration is immediate by discarding the failed node(s) and progressing forward with the SFC route without establishing a new path. Remarkably, the restored paths are shorter than the pre-fault primary paths. The methods yield noticeable restoration rates, reduced delays and costs compared to conventional methods, for single, dual-clustered, and dual-distributed failures occurring under high-congestion scenarios.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1535-1552"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148902848","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}

引用次数: 0

Coupling-Aware Edge-Cloud Computing for Multi-UAV Cooperative Sensing via Evolutionary Game-Based Deep Reinforcement Learning 基于进化博弈深度强化学习的多无人机协同感知耦合感知边缘云计算

IF 5.3 2区计算机科学

IEEE Transactions on Cloud Computing Pub Date : 2026-07-01 Epub Date: 2026-06-01 DOI: 10.1109/TCC.2026.3697781

Yishan Chen;Tong Zhou;Qing Xu;Guanjie Cheng

{"title":"Coupling-Aware Edge-Cloud Computing for Multi-UAV Cooperative Sensing via Evolutionary Game-Based Deep Reinforcement Learning","authors":"Yishan Chen;Tong Zhou;Qing Xu;Guanjie Cheng","doi":"10.1109/TCC.2026.3697781","DOIUrl":"https://doi.org/10.1109/TCC.2026.3697781","url":null,"abstract":"For multi-UAV cooperative perception applications in Intelligent Transportation Systems (ITS), edge-cloud collaborative computing is severely constrained by co-channel interference and limited edge computing resources, which induce strong network-level coupling among UAV offloading decisions. In distributed settings where global information is unavailable, this coupling exacerbates environmental non-stationarity during multi-agent training, leading to slow convergence and policy oscillations. To address these issues, we propose a distributed decision-making framework based on evolutionary game theory (EGT). Specifically, we first formulate the multi-UAV task offloading problem as an evolutionary game, capturing the payoff interdependence caused by wireless interference and edge resource contention. Subsequently, we apply replicator dynamics to analyze the evolution of offloading strategies, derive the interior evolutionary equilibrium, and establish its asymptotic stability. Finally, we develop an EGT-enhanced multi-agent deep reinforcement learning (EGT-MADQN) algorithm, which incorporates relative fitness derived from coupling-affected payoffs into strategy fusion through adaptive step-size modulation to to facilitate stable policy adaptation during training. Extensive simulation results demonstrate that the proposed method consistently outperforms baseline approaches in minimizing total cost and task latency while achieving superior convergence stability. It also exhibits strong robustness and scalability across varying UAV fleet sizes and resource configurations.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1657-1672"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148903064","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}

引用次数: 0

INTpress: Compression-Based Quantized In-Band Network Telemetry for Low Bandwidth Overhead Monitoring 基于压缩的量化带内网络遥测，用于低带宽开销监测

IF 5.3 2区计算机科学

IEEE Transactions on Cloud Computing Pub Date : 2026-07-01 Epub Date: 2026-03-06 DOI: 10.1109/TCC.2026.3690765

Chanbin Bae;Heewon Kim;Junkyu Hong;Hwimo Ku;Dongjin Lee;Sangheon Pack

{"title":"INTpress: Compression-Based Quantized In-Band Network Telemetry for Low Bandwidth Overhead Monitoring","authors":"Chanbin Bae;Heewon Kim;Junkyu Hong;Hwimo Ku;Dongjin Lee;Sangheon Pack","doi":"10.1109/TCC.2026.3690765","DOIUrl":"https://doi.org/10.1109/TCC.2026.3690765","url":null,"abstract":"Although in-band network telemetry (INT) allows for real-time and fine-grained network monitoring through a programmable data plane, it also adds to bandwidth overhead due to embedding network details in the packet headers. To mitigate this overhead, a compression technique can be used; yet, applying such compression in the data plane poses another challenge, as it may surpass available hardware resources. To tackle this, we introduce a compression-based quantized INT (INTpress), designed to efficiently decrease overhead by encoding data into shorter bit lengths in a manner native to the data plane. We also develop a dynamic data partitioning algorithm that formulates compression rules to balance monitoring accuracy and bandwidth usage. We implemented the INTpress switch using Tofino programmable switches, facilitating serialization to adeptly manage variable-length compressed bit streams, thus significantly reducing bandwidth usage. Our comprehensive experiments demonstrate the efficacy of INTpress across diverse network settings, including data center networks, backbone networks, and mobile core networks. The evaluation results show that INTpress can reduce bandwidth overhead by up to 66.01% relative to existing INT solutions, while preserving a high level of accuracy.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"1918-1932"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148903101","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}

引用次数: 0

Toward FPGA Cloud Security: A Ring Oscillator-Free and Low-Power Attack Circuit 迈向FPGA云安全：一种无环振和低功耗攻击电路

IF 5.3 2区计算机科学

IEEE Transactions on Cloud Computing Pub Date : 2026-07-01 Epub Date: 2026-03-29 DOI: 10.1109/TCC.2026.3698470

She Tang;Zhiyuan Bai;Zhe Chen;Jian Wang

{"title":"Toward FPGA Cloud Security: A Ring Oscillator-Free and Low-Power Attack Circuit","authors":"She Tang;Zhiyuan Bai;Zhe Chen;Jian Wang","doi":"10.1109/TCC.2026.3698470","DOIUrl":"https://doi.org/10.1109/TCC.2026.3698470","url":null,"abstract":"Since FPGA clouds play an important role in cutting-edge domains as computing infrastructures, their security has attracted widespread attention from security researchers. In our paper, we study a novel attack circuit that can bypass defence techniques to conduct denial-of-service (DoS) attacks on FPGA clouds, prompting researchers to enhance the defence techniques of FPGA clouds. First, we design a tree-based attack circuit that does not include ring oscillator structures and nodes with high fan-out values. Therefore, our attack circuit can elude structure detection. Then, we propose a minimal resource search algorithm based on the power and temperature constraints to reduce FPGA power and temperature by shrinking the scale of our attack circuit. With this approach, our circuit can bypass the dynamic monitoring of FPGA clouds. In order to verify the effectiveness of our attack circuit, we carry out DoS attacks on multiple types of FPGAs, including FPGAs deployed in cloud platforms. The experimental results show that our attack circuit can successfully crash these FPGAs, and pose a serious threat to FPGA clouds. Furthermore, we demonstrate that our attack circuit has the advantage of covertness compared to other DoS attack circuits.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"2223-2236"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148903115","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}

引用次数: 0

TriCooling-CAT: A Carbon Analysis Toolkit for SoC-Cluster Micro AI Data Center Design TriCooling-CAT: SoC-Cluster微AI数据中心设计的碳分析工具包

IF 5.3 2区计算机科学

IEEE Transactions on Cloud Computing Pub Date : 2026-07-01 Epub Date: 2026-07-17 DOI: 10.1109/TCC.2026.3714538

Jinyang Guo;Xinkai Wang;Jing Wang;Xiaofeng Hou;Chao Li;Minyi Guo

{"title":"TriCooling-CAT: A Carbon Analysis Toolkit for SoC-Cluster Micro AI Data Center Design","authors":"Jinyang Guo;Xinkai Wang;Jing Wang;Xiaofeng Hou;Chao Li;Minyi Guo","doi":"10.1109/TCC.2026.3714538","DOIUrl":"https://doi.org/10.1109/TCC.2026.3714538","url":null,"abstract":"Micro AI data centers (MAIDCs) integrate multiple SoC-class accelerators with local networking equipment, power delivery components, and cooling modules within a compact enclosure. In this setting, carbon efficiency is shaped not only by device energy consumption, but also by many other factors, including randomly triggered thermal throttling, unexpected execution time extension, and embodied carbon amortization, etc. Treating these factors independently can yield design choices that appear highly carbon-efficient. However, it turns out to be suboptimal once we consider their coupled effects. We refer to this issue as the Green Mirage challenge and characterize its impact on MAIDC design. We show that dynamic workload consolidation (spatial dimension), carbon emission variation (temporal dimension), and lifecycle-aware provisioning (structural dimension) can affect MAIDC carbon efficiency assessment in a complex way. We present TriCooling-CAT, a novel reduced-order carbon analysis framework for SoC-cluster MAIDC design. TriCooling-CAT allows users to make informed decisions based on joint analysis of temperature, power, cooling actuation, throttling, and execution progress. We show that TriCooling-CAT can provide useful insights for both academics and practitioners who intend to build more efficient green MAIDC systems.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"2237-2250"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148903116","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}

引用次数: 0

Optimizing Server Allocation for Multi-Tenant GPU Inference With Dynamic Batching: A Queueing Theoretic Approach 基于动态批处理的多租户GPU推理服务器分配优化：一种排队理论方法

IF 5.3 2区计算机科学

IEEE Transactions on Cloud Computing Pub Date : 2026-07-01 Epub Date: 2026-03-25 DOI: 10.1109/TCC.2026.3696613

Ryota Okuyama;Yoshiaki Inoue

{"title":"Optimizing Server Allocation for Multi-Tenant GPU Inference With Dynamic Batching: A Queueing Theoretic Approach","authors":"Ryota Okuyama;Yoshiaki Inoue","doi":"10.1109/TCC.2026.3696613","DOIUrl":"https://doi.org/10.1109/TCC.2026.3696613","url":null,"abstract":"Efficient GPU resource management is essential for multi-tenant deep neural network (DNN) inference platforms, where dynamic batching is widely used to improve throughput and energy efficiency. However, batching introduces a trade-off between latency and batch size, particularly when diverse applications compete for a limited server pool. Existing measurement-based tuning approaches often require re-profiling system performance under varying arrival rates, which renders them impractical for dynamic production environments. This paper introduces a queueing-theoretic framework for application-aware server allocation under energy constraints. We model each application’s server using a realistic dynamic batching policy and Erlang interarrival processes, arising from Poisson job arrivals and round-robin job dispatching. We derive closed-form approximations for expected batch size and job latency. These approximations enable a convex optimization that minimizes the maximum average latency across applications while satisfying batch-dependent energy efficiency constraints. The proposed method eliminates the need for exhaustive simulation or empirical profiling across different arrival rates, making it well-suited for dynamic, real-time server provisioning. Experiments on a real GPU server with ResNet-50 and RetinaNet demonstrate that our approach closely approximates optimal allocation strategies while significantly reducing computational overhead. The framework supports both server-based and serverless deployments and is readily applicable to modern DNN inference platforms.","PeriodicalId":13202,"journal":{"name":"IEEE Transactions on Cloud Computing","volume":"14 3","pages":"2017-2032"},"PeriodicalIF":5.3,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148903120","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}

引用次数: 0