Journal of Parallel and Distributed Computing最新文献_第9页

Deep embedded lightweight CNN network for indoor objects detection on FPGA 基于FPGA的室内物体检测的深度嵌入式轻量级CNN网络

IF 3.4 3区计算机科学 Q1 COMPUTER SCIENCE, THEORY & METHODS

Journal of Parallel and Distributed Computing

Pub Date : 2025-04-05 DOI: 10.1016/j.jpdc.2025.105085

Mouna Afif , Riadh Ayachi , Yahia Said , Mohamed Atri

Indoor object detection and recognition present an active research axis in computer vision and artificial intelligence fields. Various deep learning-based techniques can be applied to solve object detection problems. With the appearance of deep convolutional neural networks (DCNN) a great breakthrough for various applications was achieved. Indoor object detection presents a primary task that can assist Blind and Visually Impaired persons (BVI) during their navigation. However, building a reliable indoor object detection system used for edge device implementations still presents a serious challenge. To address this problem, we propose in this work to build an indoor object detection system based on DCNN network. Cross-stage partial network (CSPNet) was used for the detection process and a lightweight backbone based on EfficientNet v2 was used as a network backbone. To ensure a lightweight implementation of the proposed work on FPGA devices, various optimization techniques have been applied to compress the model size and reduce its computation complexity. The proposed indoor object detection system was implemented on a Xilinx ZCU 102 board. Training and testing experiments have been conducted on the proposed indoor objects dataset that counts 11,000 images containing 25 landmark classes and in indoor objects detection dataset. The proposed work achieved 82.60 mAP and 28 FPS for the original version and 80.04 with 35 FPS as processing speed for the compressed version.

室内目标检测与识别是计算机视觉和人工智能领域一个活跃的研究方向。各种基于深度学习的技术可以应用于解决目标检测问题。随着深度卷积神经网络（deep convolutional neural networks， DCNN）的出现，在各种应用上取得了很大的突破。室内目标检测是帮助盲人和视障人士（BVI）导航的主要任务。然而，建立一个可靠的室内目标检测系统用于边缘设备的实现仍然是一个严峻的挑战。针对这一问题，本文提出构建一个基于DCNN网络的室内目标检测系统。检测过程采用跨阶段局部网络（CSPNet），采用基于EfficientNet v2的轻量级骨干网作为网络骨干网。为了确保所提出的工作在FPGA器件上的轻量级实现，各种优化技术被应用于压缩模型尺寸并降低其计算复杂度。所提出的室内目标检测系统在Xilinx ZCU 102板上实现。在包含25个地标类的1.1万幅图像的室内目标数据集和室内目标检测数据集上进行了训练和测试实验。提出的工作在原始版本中实现了82.60 mAP和28 FPS，压缩版本实现了80.04和35 FPS的处理速度。

{"title":"Deep embedded lightweight CNN network for indoor objects detection on FPGA","authors":"Mouna Afif , Riadh Ayachi , Yahia Said , Mohamed Atri","doi":"10.1016/j.jpdc.2025.105085","DOIUrl":"10.1016/j.jpdc.2025.105085","url":null,"abstract":"<div><div>Indoor object detection and recognition present an active research axis in computer vision and artificial intelligence fields. Various deep learning-based techniques can be applied to solve object detection problems. With the appearance of deep convolutional neural networks (DCNN) a great breakthrough for various applications was achieved. Indoor object detection presents a primary task that can assist Blind and Visually Impaired persons (BVI) during their navigation. However, building a reliable indoor object detection system used for edge device implementations still presents a serious challenge. To address this problem, we propose in this work to build an indoor object detection system based on DCNN network. Cross-stage partial network (CSPNet) was used for the detection process and a lightweight backbone based on EfficientNet v2 was used as a network backbone. To ensure a lightweight implementation of the proposed work on FPGA devices, various optimization techniques have been applied to compress the model size and reduce its computation complexity. The proposed indoor object detection system was implemented on a Xilinx ZCU 102 board. Training and testing experiments have been conducted on the proposed indoor objects dataset that counts 11,000 images containing 25 landmark classes and in indoor objects detection dataset. The proposed work achieved 82.60 mAP and 28 FPS for the original version and 80.04 with 35 FPS as processing speed for the compressed version.</div></div>","PeriodicalId":54775,"journal":{"name":"Journal of Parallel and Distributed Computing","volume":"201 ","pages":"Article 105085"},"PeriodicalIF":3.4,"publicationDate":"2025-04-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143806913","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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

Price-aware resource management for multi-modal DNN inference in collaborative heterogeneous edge environments 协同异构边缘环境下多模态DNN推理的价格感知资源管理

IF 3.4 3区计算机科学 Q1 COMPUTER SCIENCE, THEORY & METHODS

Journal of Parallel and Distributed Computing

Pub Date : 2025-04-04 DOI: 10.1016/j.jpdc.2025.105080

Fengyi Huang , Wenhua Wang , Jianxiong Guo , Wentao Fan , Yang Xu , Tian Wang , Jiannong Cao

To address the limitations of ARM64-based AI edge devices, which are energy-efficient but computationally constrained, as well as general-purpose edge servers, this paper proposes a multi-modal CollaborativeHeterogeneous Edge Computing (CHEC) architecture that achieves low latency and enhances computational capabilities. The CHEC framework, which is segmented into an edge private cloud and an edge public cloud, endeavors to optimize the profits of Edge Service Providers (ESPs) through dynamic heterogeneous resource management. In particular, it is achieved by formulating the challenge as a multi-stage Mixed-Integer Nonlinear Programming (MINLP) problem. We introduce a resource collaboration system based on resource leasing incorporating three Economic Payment Models (EPMs), ensuring efficient and profitable resource utilization. To tackle this complex issue, we develop a three-layer Hybrid Deep Reinforcement Learning (HDRL) algorithm with EPMs, HDRL-EPMs, for efficient management of dynamic and heterogeneous resources. Extensive simulations confirm the algorithm's ability to ensure convergence and approximate optimal solutions, significantly outperforming existing methods. Testbed experiments demonstrate that the CHEC architecture reduces latency by up to 21.83% in real-world applications, markedly surpassing previous approaches.

为了解决基于arm64的人工智能边缘设备节能但计算受限以及通用边缘服务器的局限性，本文提出了一种多模态协同异构边缘计算（CHEC）架构，以实现低延迟和增强计算能力。CHEC框架分为边缘私有云和边缘公共云，通过动态异构资源管理，优化边缘服务提供商（esp）的利润。特别地，它是通过将挑战表述为多阶段混合整数非线性规划（MINLP）问题来实现的。我们引入了一个基于资源租赁的资源协作系统，该系统结合了三种经济支付模式（epm），确保了资源的高效和盈利利用。为了解决这个复杂的问题，我们开发了一种带有epm的三层混合深度强化学习（HDRL）算法，HDRL- epm，用于有效管理动态和异构资源。大量的仿真证实了该算法确保收敛和近似最优解的能力，显著优于现有方法。测试平台实验表明，CHEC架构在实际应用中减少了高达21.83%的延迟，明显优于以前的方法。

{"title":"Price-aware resource management for multi-modal DNN inference in collaborative heterogeneous edge environments","authors":"Fengyi Huang , Wenhua Wang , Jianxiong Guo , Wentao Fan , Yang Xu , Tian Wang , Jiannong Cao","doi":"10.1016/j.jpdc.2025.105080","DOIUrl":"10.1016/j.jpdc.2025.105080","url":null,"abstract":"<div><div>To address the limitations of ARM64-based AI edge devices, which are energy-efficient but computationally constrained, as well as general-purpose edge servers, this paper proposes a multi-modal CollaborativeHeterogeneous Edge Computing (CHEC) architecture that achieves low latency and enhances computational capabilities. The CHEC framework, which is segmented into an edge private cloud and an edge public cloud, endeavors to optimize the profits of Edge Service Providers (ESPs) through dynamic heterogeneous resource management. In particular, it is achieved by formulating the challenge as a multi-stage Mixed-Integer Nonlinear Programming (MINLP) problem. We introduce a resource collaboration system based on resource leasing incorporating three Economic Payment Models (EPMs), ensuring efficient and profitable resource utilization. To tackle this complex issue, we develop a three-layer Hybrid Deep Reinforcement Learning (HDRL) algorithm with EPMs, HDRL-EPMs, for efficient management of dynamic and heterogeneous resources. Extensive simulations confirm the algorithm's ability to ensure convergence and approximate optimal solutions, significantly outperforming existing methods. Testbed experiments demonstrate that the CHEC architecture reduces latency by up to 21.83% in real-world applications, markedly surpassing previous approaches.</div></div>","PeriodicalId":54775,"journal":{"name":"Journal of Parallel and Distributed Computing","volume":"201 ","pages":"Article 105080"},"PeriodicalIF":3.4,"publicationDate":"2025-04-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143792551","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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

Embedded scaffolding for teaching and assessing inquiry-based hands-on laboratory on distributed systems 嵌入式脚手架用于教学和评估基于探究的分布式系统动手实验

IF 3.4 3区计算机科学 Q1 COMPUTER SCIENCE, THEORY & METHODS

Journal of Parallel and Distributed Computing

Pub Date : 2025-04-03 DOI: 10.1016/j.jpdc.2025.105082

Jordi Guitart

Context

Information Technology education must cultivate proficiency on distributed systems, including strong hands-on laboratory skills, to meet the needs of the society and the industry. Given the complexity of distributed systems, any successful methodology to teach them to novice students must be scaffolded appropriately to ensure that the students acquire the required degree of expertise.

Objective

We propose a comprehensive scaffolding approach for inquiry-based hands-on laboratory on a distributed systems course, which guides not only the learning process, but also its assessment. The approach is based mainly on embedded scaffolds, namely explicit coding and experimental milestones and open questions with predefined grades, but also features contingent scaffolds provided by the teacher when additional assistance is needed.

Method

We apply the methodology in the context of the subject ‘Distributed Network Systems’ offered by our university. We compare the students' performance during three academic courses using the proposed methodology with respect to the three previous courses that were still using the former methodology. We use both visual representations and planned Analysis of Variance (ANOVA) tests to verify our hypothesis defined as a complex contrast.

Findings

We find that there is a statistically significant improvement in the students' performance when using the new methodology, both in their grades of the assignments (F(1, 75.364) = 17.770,

p = 6.85 \times 10^{- 5}

) and, more importantly, also in their grades of the exam questions about the practicals (F(1, 123.186) = 13.285,

p = 3.93 \times 10^{- 4}

).

Implications

Our results encourage other instructors to incorporate embedded scaffolds for teaching and assessing their hands-on laboratories on distributed systems.

背景信息技术教育必须培养对分布式系统的熟练程度，包括强大的动手实验技能，以满足社会和行业的需求。考虑到分布式系统的复杂性，任何向新手教授分布式系统的成功方法都必须有适当的框架，以确保学生获得所需的专业知识。目的在分布式系统课程中，我们提出了一种基于探究性实践实验的综合脚手架方法，该方法不仅指导了学习过程，而且指导了评估过程。该方法主要基于嵌入式框架，即明确的编码和实验里程碑以及预定义分数的开放性问题，但也有教师在需要额外帮助时提供的临时框架。方法我们将该方法应用于我校提供的“分布式网络系统”课题。我们比较了学生在三门学术课程中使用所提出的方法的表现，以及仍然使用前一种方法的三门课程。我们使用视觉表示和计划的方差分析（ANOVA）检验来验证我们定义为复杂对比的假设。我们发现，在使用新方法时，学生的表现在统计上有显着的改善，无论是在他们的作业成绩（F(1,75.364) = 17.770, p=6.85×10−5），更重要的是，在他们的实践考试问题的成绩（F(1,123.186) = 13.285, p=3.93×10−4）。我们的结果鼓励其他教师将嵌入式支架纳入分布式系统的教学和评估实践实验。

{"title":"Embedded scaffolding for teaching and assessing inquiry-based hands-on laboratory on distributed systems","authors":"Jordi Guitart","doi":"10.1016/j.jpdc.2025.105082","DOIUrl":"10.1016/j.jpdc.2025.105082","url":null,"abstract":"<div><h3>Context</h3><div>Information Technology education must cultivate proficiency on distributed systems, including strong hands-on laboratory skills, to meet the needs of the society and the industry. Given the complexity of distributed systems, any successful methodology to teach them to novice students must be scaffolded appropriately to ensure that the students acquire the required degree of expertise.</div></div><div><h3>Objective</h3><div>We propose a comprehensive scaffolding approach for inquiry-based hands-on laboratory on a distributed systems course, which guides not only the learning process, but also its assessment. The approach is based mainly on embedded scaffolds, namely explicit coding and experimental milestones and open questions with predefined grades, but also features contingent scaffolds provided by the teacher when additional assistance is needed.</div></div><div><h3>Method</h3><div>We apply the methodology in the context of the subject ‘Distributed Network Systems’ offered by our university. We compare the students' performance during three academic courses using the proposed methodology with respect to the three previous courses that were still using the former methodology. We use both visual representations and planned Analysis of Variance (ANOVA) tests to verify our hypothesis defined as a complex contrast.</div></div><div><h3>Findings</h3><div>We find that there is a statistically significant improvement in the students' performance when using the new methodology, both in their grades of the assignments (<em>F</em>(1, 75.364) = 17.770, <span><math><mi>p</mi><mo>=</mo><mn>6.85</mn><mo>×</mo><msup><mrow><mn>10</mn></mrow><mrow><mo>−</mo><mn>5</mn></mrow></msup></math></span>) and, more importantly, also in their grades of the exam questions about the practicals (<em>F</em>(1, 123.186) = 13.285, <span><math><mi>p</mi><mo>=</mo><mn>3.93</mn><mo>×</mo><msup><mrow><mn>10</mn></mrow><mrow><mo>−</mo><mn>4</mn></mrow></msup></math></span>).</div></div><div><h3>Implications</h3><div>Our results encourage other instructors to incorporate embedded scaffolds for teaching and assessing their hands-on laboratories on distributed systems.</div></div>","PeriodicalId":54775,"journal":{"name":"Journal of Parallel and Distributed Computing","volume":"201 ","pages":"Article 105082"},"PeriodicalIF":3.4,"publicationDate":"2025-04-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143792549","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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

Multi-ARCL: Multimodal adaptive relay-based distributed continual learning for encrypted traffic classification Multi-ARCL：基于多模态自适应中继的分布式持续学习，用于加密流量分类

IF 3.4 3区计算机科学 Q1 COMPUTER SCIENCE, THEORY & METHODS

Journal of Parallel and Distributed Computing

Pub Date : 2025-04-03 DOI: 10.1016/j.jpdc.2025.105083

Zeyi Li , Minyao Liu , Pan Wang , Wangyu Su , Tianshui Chang , Xuejiao Chen , Xiaokang Zhou

Encrypted Traffic Classification (ETC) using Deep Learning (DL) faces two bottlenecks: homogeneous network traffic representation and ineffective model updates. Currently, multimodal-based DL combined with the Continual Learning (CL) approaches mitigate the above problems but overlook silent applications, whose traffic is absent due to guideline violations leading developers to cease their operation and maintenance. Specifically, silent applications accelerate the decay of model stability, while new and active applications challenge model plasticity. This paper presents Multi-ARCL, a multimodal adaptive replay-based distributed CL framework for ETC. The framework prioritizes using crypto-semantic information from flows' payload and flows' statistical features to represent. Additionally, the framework proposes an adaptive relay-based continual learning method that effectively eliminates silent neurons and retrains new samples and a limited subset of old ones. Exemplars of silent applications are selectively removed during new task training. To enhance training efficiency, the framework uses distributed learning to quickly address the stability-plasticity dilemma and reduce the cost of storing silent applications. Experiments show that ARCL outperforms state-of-the-art methods, with an accuracy improvement of over 8.64% on the NJUPT2023 dataset.

基于深度学习（DL）的加密流量分类（ETC）面临两个瓶颈：同质网络流量表示和无效的模型更新。目前，基于多模式的深度学习与持续学习（CL）方法相结合缓解了上述问题，但忽略了静默应用程序，由于违反指导方针导致开发人员停止其操作和维护，其流量缺失。具体来说，沉默应用加速了模型稳定性的衰减，而新的和活跃的应用挑战了模型的可塑性。本文提出了基于多模态自适应重播的分布式CL框架Multi-ARCL。该框架优先使用来自流的有效负载和流的统计特征的加密语义信息来表示。此外，该框架提出了一种基于自适应继电器的持续学习方法，该方法有效地消除了沉默神经元，并重新训练了新样本和旧样本的有限子集。在新任务训练期间有选择地删除沉默应用程序的示例。为了提高训练效率，该框架采用分布式学习，快速解决了稳定性-可塑性难题，降低了存储静默应用程序的成本。实验表明，在NJUPT2023数据集上，ARCL的准确率提高了8.64%以上。

{"title":"Multi-ARCL: Multimodal adaptive relay-based distributed continual learning for encrypted traffic classification","authors":"Zeyi Li , Minyao Liu , Pan Wang , Wangyu Su , Tianshui Chang , Xuejiao Chen , Xiaokang Zhou","doi":"10.1016/j.jpdc.2025.105083","DOIUrl":"10.1016/j.jpdc.2025.105083","url":null,"abstract":"<div><div>Encrypted Traffic Classification (ETC) using Deep Learning (DL) faces two bottlenecks: homogeneous network traffic representation and ineffective model updates. Currently, multimodal-based DL combined with the Continual Learning (CL) approaches mitigate the above problems but overlook silent applications, whose traffic is absent due to guideline violations leading developers to cease their operation and maintenance. Specifically, silent applications accelerate the decay of model stability, while new and active applications challenge model plasticity. This paper presents Multi-ARCL, a multimodal adaptive replay-based distributed CL framework for ETC. The framework prioritizes using crypto-semantic information from flows' payload and flows' statistical features to represent. Additionally, the framework proposes an adaptive relay-based continual learning method that effectively eliminates silent neurons and retrains new samples and a limited subset of old ones. Exemplars of silent applications are selectively removed during new task training. To enhance training efficiency, the framework uses distributed learning to quickly address the stability-plasticity dilemma and reduce the cost of storing silent applications. Experiments show that ARCL outperforms state-of-the-art methods, with an accuracy improvement of over 8.64% on the NJUPT2023 dataset.</div></div>","PeriodicalId":54775,"journal":{"name":"Journal of Parallel and Distributed Computing","volume":"201 ","pages":"Article 105083"},"PeriodicalIF":3.4,"publicationDate":"2025-04-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143800423","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

The European master for HPC curriculum 欧洲HPC硕士课程

IF 3.4 3区计算机科学 Q1 COMPUTER SCIENCE, THEORY & METHODS

Journal of Parallel and Distributed Computing

Pub Date : 2025-04-03 DOI: 10.1016/j.jpdc.2025.105081

Pascal Bouvry , Mats Brorsson , Ramon Canal , Aryan Eftekhari , Siegfried Höfinger , Didier Smets , Harald Köstler , Tomáš Kozubek , Ezhilmathi Krishnasamy , Josep Llosa , Alexandra Lukas-Rother , Xavier Martorell , Dirk Pleiter , Ana Proykova , Maria-Ribera Sancho , Olaf Schenk , Cristina Silvano

The use of High-Performance Computing (HPC) is crucial for addressing various grand challenges. While significant investments are made in digital infrastructures that comprise HPC resources, its realisation, operation, and, in particular, its use critically depends on suitably trained experts. In this paper, we present the results of an effort to design and implement a pan-European reference curriculum for a master's degree in HPC.

高性能计算（HPC）的使用对于解决各种重大挑战至关重要。虽然在包含高性能计算资源的数字基础设施上进行了大量投资，但它的实现、运营，特别是它的使用，在很大程度上取决于受过适当培训的专家。在本文中，我们展示了设计和实施泛欧HPC硕士学位参考课程的成果。

引用次数: 0

STVAI: Exploring spatio-temporal similarity for scalable and efficient intelligent video inference STVAI：探索可扩展和高效智能视频推理的时空相似性

IF 3.4 3区计算机科学 Q1 COMPUTER SCIENCE, THEORY & METHODS

Journal of Parallel and Distributed Computing

Pub Date : 2025-04-03 DOI: 10.1016/j.jpdc.2025.105079

Chuang Li , Heshi Wang , Yanhua Wen , Qingyu Shi , Qinyu Wang , Chunhua Hu , Dongchen Wu

The integration of video data computation and inference is a cornerstone for the evolution of multimodal artificial intelligence (MAI). The extensive adoption and optimization of CNN-based frameworks have significantly improved the accuracy of video inference, yet they present substantial challenges for real-time and large-scale computational demands. Existing researches primarily utilize the temporal similarity between video frames to reduce redundant computations, but most of them overlooked the spatial similarity within the frames themselves. Hence, we propose STVAI, a scalable and efficient method that leverages both spatial and temporal similarities to accelerate video inference. This approach uses a parallel region merging strategy, which maintains inference accuracy and enhances the sparsity of the computation matrix. Moreover, we have optimized the computation of sparse convolutions by utilizing Tensor Cores, which accelerate dense convolution computations based on the sparsity of the tiles. Experimental results demonstrate that STVAI achieves a stable acceleration of 1.25 times faster than cuDNN implementations, with only a 5% decrease in prediction accuracy. STVAI can achieve accelerations up to 1.53x, surpassing that of existing methods. Our method can be directly applied to various CNN architectures for video inference tasks without the need for retraining the model.

视频数据计算和推理的集成是多模态人工智能（MAI）发展的基石。基于cnn的框架的广泛采用和优化大大提高了视频推理的准确性，但它们对实时和大规模计算需求提出了实质性挑战。现有研究主要利用视频帧之间的时间相似性来减少冗余计算，但大多忽略了帧本身的空间相似性。因此，我们提出了STVAI，这是一种可扩展且高效的方法，它利用空间和时间相似性来加速视频推理。该方法采用并行区域合并策略，既保持了推理精度，又提高了计算矩阵的稀疏性。此外，我们还利用Tensor Cores优化了稀疏卷积的计算，该算法基于贴图的稀疏性加速了密集卷积的计算。实验结果表明，STVAI实现的稳定加速速度比cuDNN实现快1.25倍，预测精度仅下降5%。STVAI可以实现高达1.53倍的加速度，超过了现有的方法。我们的方法可以直接应用于各种CNN架构的视频推理任务，而不需要对模型进行重新训练。

{"title":"STVAI: Exploring spatio-temporal similarity for scalable and efficient intelligent video inference","authors":"Chuang Li , Heshi Wang , Yanhua Wen , Qingyu Shi , Qinyu Wang , Chunhua Hu , Dongchen Wu","doi":"10.1016/j.jpdc.2025.105079","DOIUrl":"10.1016/j.jpdc.2025.105079","url":null,"abstract":"<div><div>The integration of video data computation and inference is a cornerstone for the evolution of multimodal artificial intelligence (MAI). The extensive adoption and optimization of CNN-based frameworks have significantly improved the accuracy of video inference, yet they present substantial challenges for real-time and large-scale computational demands. Existing researches primarily utilize the temporal similarity between video frames to reduce redundant computations, but most of them overlooked the spatial similarity within the frames themselves. Hence, we propose STVAI, a scalable and efficient method that leverages both spatial and temporal similarities to accelerate video inference. This approach uses a parallel region merging strategy, which maintains inference accuracy and enhances the sparsity of the computation matrix. Moreover, we have optimized the computation of sparse convolutions by utilizing Tensor Cores, which accelerate dense convolution computations based on the sparsity of the tiles. Experimental results demonstrate that STVAI achieves a stable acceleration of 1.25 times faster than cuDNN implementations, with only a 5% decrease in prediction accuracy. STVAI can achieve accelerations up to 1.53x, surpassing that of existing methods. Our method can be directly applied to various CNN architectures for video inference tasks without the need for retraining the model.</div></div>","PeriodicalId":54775,"journal":{"name":"Journal of Parallel and Distributed Computing","volume":"201 ","pages":"Article 105079"},"PeriodicalIF":3.4,"publicationDate":"2025-04-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143792552","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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

MMBypass: Towards efficient multi-modal AI computing with adaptive bypass network MMBypass：利用自适应旁路网络实现高效的多模式人工智能计算

IF 3.4 3区计算机科学 Q1 COMPUTER SCIENCE, THEORY & METHODS

Journal of Parallel and Distributed Computing

Pub Date : 2025-04-03 DOI: 10.1016/j.jpdc.2025.105078

Yifei Pu , Xinfeng Xia , Xiaofeng Hou , Chi Wang , Cheng Xu , Jiacheng Liu , Jing Wang , Minyi Guo , Jingling Yuan , Chao Li

Multi-modal artificial intelligence systems demonstrate superior performance through cross-modal information fusion and processing mechanisms, surpassing conventional unimodal architectures. However, the enhanced computational complexity required for processing heterogeneous data streams in multi-modal frameworks results in elevated inference latency compared to their uni-modal architectures. This limitation significantly constrains deployment feasibility for real-time and large-scale applications. To address this challenge, we present MMBypass, an adaptive and efficient architecture for multi-modal AI acceleration. Our solution implements intelligent layer-skipping mechanisms through adaptive computational complexity analysis of multi-modal tasks, achieving latency reduction while maintaining predictive accuracy and mitigating model overfitting in specialized scenarios. The architecture's innovation lies in two aspects: 1) We design bypasses for each uni-modal network in multi-modal networks to perform adaptive computing. 2) We design a guider to dynamically choose the optimal bypasses. Distinct from existing methods, MMBypass maintains broad applicability without requiring domain-specific prerequisites, and it shows significantly better performance on data samples with different difficulties. Empirical evaluations demonstrate our architecture achieves 44.5% average latency reduction while matching or exceeding baseline accuracy across diverse multi-modal benchmarks.

多模态人工智能系统通过跨模态信息融合和处理机制，超越了传统的单模态架构，展现出卓越的性能。然而，与单模态架构相比，在多模态框架中处理异构数据流所需的计算复杂性增加了推理延迟。这一限制极大地限制了实时和大规模应用程序的部署可行性。为了应对这一挑战，我们提出了MMBypass，这是一种用于多模态人工智能加速的自适应高效架构。我们的解决方案通过自适应多模式任务的计算复杂性分析实现智能跳层机制，在保持预测准确性的同时减少延迟，并在特定场景中减轻模型过拟合。该体系结构的创新之处在于两个方面：1)在多模态网络中，我们为每个单模态网络设计旁路来进行自适应计算。2)设计了一个导流器来动态选择最优旁路。与现有方法不同，MMBypass方法保持了广泛的适用性，不需要特定领域的先决条件，并且在不同难度的数据样本上表现出明显更好的性能。经验评估表明，我们的架构实现了44.5%的平均延迟减少，同时在不同的多模式基准测试中达到或超过基线精度。

{"title":"MMBypass: Towards efficient multi-modal AI computing with adaptive bypass network","authors":"Yifei Pu , Xinfeng Xia , Xiaofeng Hou , Chi Wang , Cheng Xu , Jiacheng Liu , Jing Wang , Minyi Guo , Jingling Yuan , Chao Li","doi":"10.1016/j.jpdc.2025.105078","DOIUrl":"10.1016/j.jpdc.2025.105078","url":null,"abstract":"<div><div>Multi-modal artificial intelligence systems demonstrate superior performance through cross-modal information fusion and processing mechanisms, surpassing conventional unimodal architectures. However, the enhanced computational complexity required for processing heterogeneous data streams in multi-modal frameworks results in elevated inference latency compared to their uni-modal architectures. This limitation significantly constrains deployment feasibility for real-time and large-scale applications. To address this challenge, we present <em>MMBypass</em>, an adaptive and efficient architecture for multi-modal AI acceleration. Our solution implements intelligent layer-skipping mechanisms through adaptive computational complexity analysis of multi-modal tasks, achieving latency reduction while maintaining predictive accuracy and mitigating model overfitting in specialized scenarios. The architecture's innovation lies in two aspects: 1) We design bypasses for each uni-modal network in multi-modal networks to perform adaptive computing. 2) We design a guider to dynamically choose the optimal bypasses. Distinct from existing methods, <em>MMBypass</em> maintains broad applicability without requiring domain-specific prerequisites, and it shows significantly better performance on data samples with different difficulties. Empirical evaluations demonstrate our architecture achieves 44.5% average latency reduction while matching or exceeding baseline accuracy across diverse multi-modal benchmarks.</div></div>","PeriodicalId":54775,"journal":{"name":"Journal of Parallel and Distributed Computing","volume":"201 ","pages":"Article 105078"},"PeriodicalIF":3.4,"publicationDate":"2025-04-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143825455","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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

Design of energy-aware sensor networks for climate and pollution monitoring 用于气候和污染监测的能量感知传感器网络设计

IF 3.4 3区计算机科学 Q1 COMPUTER SCIENCE, THEORY & METHODS

Journal of Parallel and Distributed Computing

Pub Date : 2025-04-02 DOI: 10.1016/j.jpdc.2025.105084

Meeniga Vijaya Lakshmi , M. Sri Raghavendra , MaddalaVijaya Lakshmi

The growing concern over climate change and Pollution has driven the development of energy-efficient sensor networks for environmental monitoring. This research proposes an energy-aware sensor network using Spanning Tree-Reinforcement Learning (ST-RL) to optimize data accuracy, minimize energy consumption, and extend the network's lifetime. The proposed method achieves significant performance improvements compared to existing approaches. Experimental results demonstrate that ST-RL enhances network lifetime by 28.57 %, reduces energy consumption by 41.24 %, improves packet delivery ratio by 3.7 %, and reduces transmission delay by 10 % over traditional methods such as EDAL, FT-EEC, and EAEDAR. The data is collected from multiple environmental sensors, processed using spanning tree algorithms for optimized connectivity and refined with reinforcement learning to suppress unnecessary transmissions. The results confirm that the proposed ST-RL technique significantly enhances energy efficiency and network reliability, making it a promising solution for large-scale climate and pollution monitoring applications.

人们对气候变化和污染的日益关注推动了用于环境监测的高能效传感器网络的发展。本研究提出了一种使用生成树-强化学习（ST-RL）的能量感知传感器网络，以优化数据准确性、最小化能耗并延长网络寿命。与现有方法相比，所提出的方法能显著提高性能。实验结果表明，与 EDAL、FT-EEC 和 EAEDAR 等传统方法相比，ST-RL 可将网络寿命延长 28.57%，能耗降低 41.24%，数据包传送率提高 3.7%，传输延迟降低 10%。数据收集自多个环境传感器，使用生成树算法进行处理以优化连接性，并通过强化学习来抑制不必要的传输。研究结果证实，所提出的 ST-RL 技术能显著提高能源效率和网络可靠性，是大规模气候和污染监测应用的理想解决方案。

{"title":"Design of energy-aware sensor networks for climate and pollution monitoring","authors":"Meeniga Vijaya Lakshmi , M. Sri Raghavendra , MaddalaVijaya Lakshmi","doi":"10.1016/j.jpdc.2025.105084","DOIUrl":"10.1016/j.jpdc.2025.105084","url":null,"abstract":"<div><div>The growing concern over climate change and Pollution has driven the development of energy-efficient sensor networks for environmental monitoring. This research proposes an energy-aware sensor network using Spanning Tree-Reinforcement Learning (ST-RL) to optimize data accuracy, minimize energy consumption, and extend the network's lifetime. The proposed method achieves significant performance improvements compared to existing approaches. Experimental results demonstrate that ST-RL enhances network lifetime by 28.57 %, reduces energy consumption by 41.24 %, improves packet delivery ratio by 3.7 %, and reduces transmission delay by 10 % over traditional methods such as EDAL, FT-EEC, and EAEDAR. The data is collected from multiple environmental sensors, processed using spanning tree algorithms for optimized connectivity and refined with reinforcement learning to suppress unnecessary transmissions. The results confirm that the proposed ST-RL technique significantly enhances energy efficiency and network reliability, making it a promising solution for large-scale climate and pollution monitoring applications.</div></div>","PeriodicalId":54775,"journal":{"name":"Journal of Parallel and Distributed Computing","volume":"201 ","pages":"Article 105084"},"PeriodicalIF":3.4,"publicationDate":"2025-04-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143824009","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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

Latency and cost-aware consumer group autoscaling in message broker systems 消息代理系统中的延迟和成本敏感消费者组自动伸缩

IF 3.4 3区计算机科学 Q1 COMPUTER SCIENCE, THEORY & METHODS

Journal of Parallel and Distributed Computing

Pub Date : 2025-03-28 DOI: 10.1016/j.jpdc.2025.105071

Diogo Landau , Nishant Saurabh , Xavier Andrade , Jorge G. Barbosa

Message brokers often facilitate communication between data producers and consumers by adding variable-sized messages to ordered distributed queues. Our goal is to determine the number of consumers and consumer partition assignments needed to ensure that the data consumption rate matches the data production rate. We model this problem as a variable item size bin packing problem. As the production rate varies, new consumer–partition assignments are computed, potentially requiring the reallocation of partitions from one consumer to another. During reallocation, data in the queue are not read, leading to increased latency costs. To address this problem, we focus on the multiobjective optimization cost of minimizing the number of consumers and reducing latency. We introduce several heuristic algorithms and compare them to state-of-the-art heuristics. In our experimental setup, the proposed modified worst fit (MWF) heuristic achieves a 48% reduction, with a similar number of consumers, in comparison with the best fit decrease (BFD). In addition, MWF achieves a

99^{t h}

percentile latency of 2.24 seconds compared with that of 364.66 with the approach by Kafka using the same number of consumers. Alternatively, to obtain a lower

99^{t h}

percentile latency than our approach does, Kafka requires at least 60% more consumers than our method requires.

消息代理通常通过向有序的分布式队列中添加可变大小的消息来促进数据生产者和消费者之间的通信。我们的目标是确定确保数据消耗率与数据产生率匹配所需的消费者数量和消费者分区分配。我们将此问题建模为可变物品大小的装箱问题。随着生产速率的变化，计算新的消费者-分区分配，可能需要将分区从一个消费者重新分配到另一个消费者。在重新分配期间，不会读取队列中的数据，从而导致延迟成本增加。为了解决这个问题，我们将重点放在最小化消费者数量和减少延迟的多目标优化成本上。我们介绍了几种启发式算法，并将它们与最先进的启发式算法进行比较。在我们的实验设置中，与最佳拟合减少（BFD）相比，所提出的改进的最坏拟合（MWF）启发式在消费者数量相似的情况下实现了48%的减少。此外，在使用相同数量的消费者时，MWF实现了2.24秒的99百分位延迟，而Kafka的方法为364.66秒。或者，为了获得比我们的方法更低的99百分位延迟，Kafka需要比我们的方法至少多60%的消费者。

{"title":"Latency and cost-aware consumer group autoscaling in message broker systems","authors":"Diogo Landau , Nishant Saurabh , Xavier Andrade , Jorge G. Barbosa","doi":"10.1016/j.jpdc.2025.105071","DOIUrl":"10.1016/j.jpdc.2025.105071","url":null,"abstract":"<div><div>Message brokers often facilitate communication between data producers and consumers by adding variable-sized messages to ordered distributed queues. Our goal is to determine the number of consumers and consumer partition assignments needed to ensure that the data consumption rate matches the data production rate. We model this problem as a variable item size bin packing problem. As the production rate varies, new consumer–partition assignments are computed, potentially requiring the reallocation of partitions from one consumer to another. During reallocation, data in the queue are not read, leading to increased latency costs. To address this problem, we focus on the multiobjective optimization cost of minimizing the number of consumers and reducing latency. We introduce several heuristic algorithms and compare them to state-of-the-art heuristics. In our experimental setup, the proposed modified worst fit (MWF) heuristic achieves a 48% reduction, with a similar number of consumers, in comparison with the best fit decrease (BFD). In addition, MWF achieves a <span><math><msup><mrow><mn>99</mn></mrow><mrow><mi>t</mi><mi>h</mi></mrow></msup></math></span> percentile latency of 2.24 seconds compared with that of 364.66 with the approach by Kafka using the same number of consumers. Alternatively, to obtain a lower <span><math><msup><mrow><mn>99</mn></mrow><mrow><mi>t</mi><mi>h</mi></mrow></msup></math></span> percentile latency than our approach does, Kafka requires at least 60% more consumers than our method requires.</div></div>","PeriodicalId":54775,"journal":{"name":"Journal of Parallel and Distributed Computing","volume":"201 ","pages":"Article 105071"},"PeriodicalIF":3.4,"publicationDate":"2025-03-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143748704","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Optimizing the layout of embedding BCube into grid architectures 优化BCube嵌入网格架构的布局

IF 3.4 3区计算机科学 Q1 COMPUTER SCIENCE, THEORY & METHODS

Journal of Parallel and Distributed Computing

Pub Date : 2025-03-27 DOI: 10.1016/j.jpdc.2025.105070

Paul Immanuel, A. Berin Greeni

The storage, processing, and distribution of enormous volumes of data are made possible by data centers, which are vital components of the contemporary computing infrastructure. The BCube network is a type of significant data center network which was developed for modular data centers that are based on shipping containers. Network embedding of data centers into certain topologies offers several benefits, including improved scalability, reduced power consumption, enhanced reliability, and improved overall network performance. Embedding of a guest graph into suitable host graphs have significant applications like: virtualizing the Network-on-Chip layouts, portability of algorithms, and the simulation capabilities of parallel architectures. A crucial key factor that influences the quality of embedding is layout. So far, there have been few results regarding the embedding of graphs into certain data center networks. However, these results are obtained by fixing data center networks as host graphs with linear arrays and cycles as guest graphs. In this work, we investigate the edge isoperimetric features of BCube and embed it into linear arrays and grid structures by considering it as the guest graph. This study is the first that we are aware of, on embedding data center networks for minimum layout.

数据中心是当代计算基础设施的重要组成部分，它使海量数据的存储、处理和分发成为可能。BCube网络是一种重要的数据中心网络，它是为基于集装箱的模块化数据中心而开发的。将数据中心的网络嵌入到某些拓扑结构中有几个好处，包括改进的可伸缩性、降低的功耗、增强的可靠性和改进的整体网络性能。将来宾图嵌入到适当的主机图中具有重要的应用，例如：虚拟化片上网络布局、算法的可移植性和并行体系结构的模拟功能。影响嵌入质量的一个关键因素是布局。到目前为止，关于将图形嵌入到某些数据中心网络中还没有什么结果。然而，这些结果是通过将数据中心网络固定为主机图，将线性阵列和周期固定为访客图来获得的。在这项工作中，我们研究了BCube的边缘等周特征，并将其作为来宾图嵌入到线性阵列和网格结构中。这是我们所知的第一个关于嵌入数据中心网络以实现最小布局的研究。

{"title":"Optimizing the layout of embedding BCube into grid architectures","authors":"Paul Immanuel, A. Berin Greeni","doi":"10.1016/j.jpdc.2025.105070","DOIUrl":"10.1016/j.jpdc.2025.105070","url":null,"abstract":"<div><div>The storage, processing, and distribution of enormous volumes of data are made possible by data centers, which are vital components of the contemporary computing infrastructure. The BCube network is a type of significant data center network which was developed for modular data centers that are based on shipping containers. Network embedding of data centers into certain topologies offers several benefits, including improved scalability, reduced power consumption, enhanced reliability, and improved overall network performance. Embedding of a guest graph into suitable host graphs have significant applications like: virtualizing the Network-on-Chip layouts, portability of algorithms, and the simulation capabilities of parallel architectures. A crucial key factor that influences the quality of embedding is layout. So far, there have been few results regarding the embedding of graphs into certain data center networks. However, these results are obtained by fixing data center networks as host graphs with linear arrays and cycles as guest graphs. In this work, we investigate the edge isoperimetric features of BCube and embed it into linear arrays and grid structures by considering it as the guest graph. This study is the first that we are aware of, on embedding data center networks for minimum layout.</div></div>","PeriodicalId":54775,"journal":{"name":"Journal of Parallel and Distributed Computing","volume":"201 ","pages":"Article 105070"},"PeriodicalIF":3.4,"publicationDate":"2025-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143738837","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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