检索结果-内蒙古大学图书馆

DeeperThings: Fully distributed CNN Inference on Resource-Constrained Edge Devices

INTERNATIONAL JOURNAL OF PARALLEL PROGRAMMING 2021年第4期49卷 600-624页

作者： Stahl, Rafael Hoffman, Alexander Mueller-Gritschneder, Daniel Gerstlauer, Andreas Schlichtmann, Ulf Tech Univ Munich Munich Germany Univ Texas Austin Austin TX 78712 USA

Performing inference of Convolutional Neural Networks (CNNs) on Internet of Things (IoT) edge devices ensures both privacy of input data and possible run time reductions when compared to a cloud solution. As most edge devices are memory- and compute-constrained, they cannot store and execute complex CNNs. Partitioning and distributing layer information across multiple edge devices to reduce the amount of computation and data on each device presents a solution to this problem. In this article, we propose DeeperThings, an approach that supports a full distribution of CNN inference tasks by partitioning fully-connected as well as both feature- and weight-intensive convolutional layers. Additionally, we jointly optimize memory, computation and communication demands. This is achieved using techniques to combine both feature and weight partitioning with a communication-aware layer fusion method, enabling holistic optimization across layers. For a given number of edge devices, the schemes are applied jointly using Integer Linear Programming (ILP) formulations to minimize data exchanged between devices, to optimize run times and to find the entire model's minimal memory footprint. Experimental results from a real-world hardware setup running four different CNN models confirm that the scheme is able to evenly balance the memory footprint between devices. For six devices on 100 Mbit/s connections the integration of layer fusion additionally leads to a reduction of communication demands by up to 28.8%. This results in run time speed-up of the inference task by up to 1.52x compared to layer partitioning without fusing.

关键词： Deep learning distributed computing IoT

来源：评论

学校读者我要写书评

暂无评论

A Data-Aware Scheduling Strategy for Executing Large-Scale distributed Workflows

引用

IEEE ACCESS 2021年 9卷 47354-47364页

作者： Giampa, Salvatore Belcastro, Loris Marozzo, Fabrizio Talia, Domenico Trunfio, Paolo Univ Calabria DIMES I-87036 Arcavacata Di Rende Italy Univ Calabria DtoK Lab Srl I-87036 Arcavacata Di Rende Italy

Task scheduling is a crucial key component for the efficient execution of data-intensive applications on distributed environments, by which many machines must be coordinated to reduce execution times and bandwidth consumption. This paper presents ADAGE, a data-aware scheduler designed to efficiently execute data-intensive workflows in large-scale computers. The proposed scheduler is based on three key features: i) critical path analysis, for discovering the critical tasks of a workflow and reducing data transferring between nodes;ii) work giving, a new dynamic planning strategy for migrating tasks from overloaded to unloaded nodes;and iii) task replication, which executes task replicas on different nodes for improving both execution time and fault tolerance. Experiments performed on a distributed computing environment composed of up to 1,024 processing nodes show that ADAGE achieves better performances than existing scheduling systems, obtaining an average reduction of up to 66% in execution time.

关键词： Data-aware scheduler workflow scheduling distributed workflows parallel programming distributed computing exascale computing

来源：评论

学校读者我要写书评

暂无评论

Codes Trading Upload for Download Cost in Secure distributed Matrix Multiplication

引用

IEEE TRANSACTIONS ON COMMUNICATIONS 2021年第8期69卷 5409-5424页

作者： Kakar, Jaber Khristoforov, Anton Ebadifar, Seyedhamed Sezgin, Aydin Ruhr Univ Bochum Inst Digital Commun Syst D-44801 Bochum Germany

In secure distributed matrix multiplication (SDMM) the multiplication AB from two private matrices A and B is outsourced to N distributed servers. In l-SDMM, the goal is to design a joint communication-computation procedure that optimally balances conflicting communication and computation metrics without leaking any information on both A and B to any set of l(i) and (B) over tilde (i) representing encoded versions of A and B destined to the i-th server. Now, SDMM involves multiple tradeoffs. One such tradeoff is the tradeoff between upload (UL) and download (DL) costs. To find a good balance between these two metrics, we propose two schemes which we term USCSA and GSCSA that are based on secure cross subspace alignment (SCSA). We show that there are various scenarios where they outperform existing SDMM schemes from the literature with respect to UL-DL efficiency. Next, we implement schemes from the literature, including USCSA and GSCSA, and test their performance on Amazon EC2. Our numerical results show that USCSA and GSCSA establish a good balance between the time spent on the communication and computation in SDMMs. This is because they combine advantages of polynomial codes, namely low time for the upload of (A) over tilde (i) and (B) over tilde (i), i is an element of [1 : N], and the computation of O-i = (A) over tilde (i) (B) over tilde (i), with those of SCSA, being a low timing overhead for the download of O-i, i is an element of [1 : N], and the decoding of AB.

关键词： Matrix multiplication security interference alignment distributed computing

来源：评论

学校读者我要写书评

暂无评论

An Incremental Iterative Acceleration Architecture in distributed Heterogeneous Environments With GPUs for Deep Learning

引用

IEEE TRANSACTIONS ON PARALLEL AND distributed SYSTEMS 2021年第11期32卷 2823-2837页

作者： Zhang, Xuedong Tang, Zhuo Du, Lifan Yang, Li Hunan Univ Coll Comp Sci & Elect Engn Natl Supercomp Ctr Changsha Changsha 410082 Peoples R China Natl Univ Def Technol Sci & Technol Parallel & Distributed Proc Lab Changsha 410073 Peoples R China Changsha Univ Sci & Technol Coll Comp & Commun Engn Changsha 410076 Hunan Peoples R China

The parallel computing capabilities of GPUs have a significant impact on computationally intensive iterative tasks. Offloading part or all of the deep learning tasks from the CPU to the GPU for execution is mainstream. However, a large number of redundant iterative calculations exist in the iterative process of computing tasks. Therefore, we propose a GPU-based distributed incremental iterative computing architecture that can make full use of distributed parallel computing and GPU memory structure. The architecture supports deep learning and other computationally intensive iterative applications by optimizing data placement and reducing redundant iterative calculations. To support block-based data partitioning and coalesced memory access on GPUs, we propose GDataSet, an abstract data set. The GPU incremental iteration manager called GTracker is designed to be responsible for GDataSet cache management on the GPU. In order to solve the limitation of on-chip memory size, we propose a variable sliding window mechanism. It improves the hit rate of cache access and the speed of data access by realizing the best block arrangement between on-chip memory and off-chip memory. Besides, a communication channel based on an incremental iterative model is designed to support data transmission and task communication in cluster computing. Finally, we implement the proposed architecture based on Spark 2.4.1 and CUDA 10.0. Comparative experiments with widely used computationally intensive iterative applications (K-means, LSTM, etc.) show that the incremental iterative acceleration architecture can significantly improve the efficiency of iterative computing.

关键词： Graphics processing units Computational modeling Deep learning Training Task analysis Sparks Parallel processing Deep learning distributed computing incremental iteration GPU

来源：评论

学校读者我要写书评

暂无评论

Collaborative Inference Acceleration with Non-Penetrative Tensor Partitioning

Collaborative Inference Acceleration with Non-Penetrative Te...

引用

2025 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2025

作者： Liu, Zhibang Xu, Chaonong Lv, Zhenjie Liu, Zhizhuo Zhao, Suyu Beijing Key Lab of Petroleum Data Mining China University of Petroleum Beijing China

ISBN: (纸本)9798350368741

The inference of large-sized images on Internet of Things (IoT) devices is commonly hindered by limited resources, while there are often stringent latency requirements for Deep Neural Network (DNN) inference. Currently, this problem is generally addressed by collaborative inference, where the large-sized image is partitioned into multiple tiles, and each tile is assigned to an IoT device for processing. However, since significant latency will be incurred due to the communication overhead caused by tile sharing, the existing collaborative inference strategy is inefficient for convolutional computation, which is indispensable for any DNN. To reduce it, we propose Non-Penetrative Tensor Partitioning (NPTP), a fine-grained tensor partitioning method that reduces the communication latency by minimizing the communication load of tiles shared, thereby reducing inference latency. We evaluate NPTP with four widely-adopted DNN models. Experimental results demonstrate that NPTP achieves a 1.44-1.68× inference speedup relative to CoEdge, a state-of-the-art (SOTA) collaborative inference algorithm. © 2025 IEEE.

关键词： collaborative inference distributed computing edge intelligence

来源：评论

学校读者我要写书评

暂无评论

Security-driven distributed platforms for intellectual property resource provision-a case study of TSITE IP

引用

JOURNAL OF CLOUD computing-ADVANCES SYSTEMS AND APPLICATIONS 2021年第1期10卷 1-15页

作者： Wang, Yang Nanjing Univ Sci & Technol Sch Intellectual Property Nanjing Peoples R China

In the information age, with the vigorous development of big data and artificial intelligence, intellectual property protection is an essential part of the current scientific and technological development. Intellectual property related data grows in a geometric progression, so the demand for IP data storage space is also increasing day by day. With the rise of cloud computing technology, intellectual property data distributed platforms based on cloud storage have also been produced one after another. Because the biggest feature of cloud storage is that storage is a service, it puts forward higher requirements for the intellectual property service industry. Firstly, it introduces the domestic intellectual property cloud platform services from the perspectives of government support, state-owned enterprises and private enterprises;Secondly, four typical distributed platforms provided by commercial resources are selected to introduce their operation modes, focusing on the problems faced by domestic intellectual property service modes;Secondly, it compares and discusses the current situation of domestic intellectual property distributed platforms;Then, aiming at the current domestic intellectual property service mode, taking tsite as an example, this paper puts forward the design and construction strategy of intellectual property protection, intellectual property operation service distributed platform and operation service mode under the background of information age.

关键词： Intellectual property Cyber security Privacy protection Operation service distributed computing TSITE

来源：评论

学校读者我要写书评

暂无评论

An efficient distributed approach for a self-healing smart grid using minimal spanning tree

引用

INTERNATIONAL JOURNAL OF ENERGY RESEARCH 2021年第10期45卷 15049-15084页

作者： Nedjah, Nadia Cardoso, Kleber Hochwart Mourelle, Luiza Macedo Univ Estado Rio De Janeiro Dept Elect Engn & Telecommun Rio De Janeiro Brazil Univ Estado Rio De Janeiro Dept Syst Engn & Computat Rio De Janeiro Brazil Rua Sao Francisco Xavier 524 Sala 5145-F BR-20550013 Rio De Janeiro Maracana Brazil

In the event of a network failure that compromises energy supply, the characteristic of smart grid self-healing consists of finding a proposal for reconfiguration of the grid, aiming at restoring the power, partially or completely to supply all network nodes. The search for a satisfactory reconfiguration is a combinatorial problem whose complexity is proportional to the network size. An exhaustive search-based method is time-consuming and often computationally non-viable. To overcome this difficulty, techniques for generating minimal spanning trees (MSTs) of the graph that represents the smart grid, are exploited. However, existing studies provide centralized implementations. In this work, we propose a distributed implementation, where each of the network switch collaborates in developing of the recovery solution. The proposed decentralized approach seeks a reduction of reconfiguration time requirements, thus increasing network intelligence. For this purpose, a distributed algorithm for building the MST is embedded in the processing elements available at the commutation nodes. The evaluated case studies show that, whenever possible, the proposed solution allows for a successful reconfiguration, regardless of the number of simultaneous failures. Moreover, the network reconfiguration time is not significantly impacted by the number of buses and included lines. The implementation presents results of communication cost and reconfiguration time significantly lower than the expected upper bounds. Notably, for the case studies, the proposed implementation achieves 70% and 69% speedup regarding the reconfiguration time from simple and multiple failures, respectively, when compared to the expected theoretical performance. Furthermore, when compared to existing multiagent-based self-healing systems, the proposed implementation recovers from failures twice as fast, making it more desirable in a smart-grid real implementation.

关键词： distributed computing minimum spanning tree self‐ healing smart grids

来源：评论

学校读者我要写书评

暂无评论

Scheduling Big Machine Learning Tasks on Clusters of Heterogeneous Edge Devices 17

Scheduling Big Machine Learning Tasks on Clusters of Heterog...

引用

17th International Conference on COMmunication Systems and NETworkS, COMSNETS 2025

作者： Mathkar, Saumya Aiyer, Shreya Bapat, Yashovardhan Pinki, Pinki Paul, Arnab K. Naik, Vinayak BITS Pilani Department of CSIS Goa India

ISBN: (纸本)9798331531195

Edge computing has transformed machine learning by using computing closer to the data sources, thereby reducing latency. The ever-increasing volume of data has necessitated forming clusters of edge devices, possibly with heterogeneous capabilities. Managing heterogeneous resources such as computation and memory remains challenging. Given the capabilities of the edge devices, we need a simple technique suitable for an Edge computing *** introduce a scheduling mechanism that leverages Integer Linear Programming (ILP) to optimize the overall computation time of ML-based tasks. We implement our scheduling mechanism that efficiently allocates resources, ensuring tasks are executed parallel across the cores of edge devices in a cluster by minimizing computation time. For tasks large enough to fit on any core, we leverage distributed learning to train the model in pieces and later combine them. We employ our ILP-based scheduler for efficient task allocation and compare its performance with a Greedy, simple approach based on best-fit technique. We evaluate our approach on three cases of sets of tasks on the spectrum of uniformity and size. Our results demonstrate two times the speed gain for our ILP-based approach over Greedy for the category with the least uniformity and large size. © 2025 IEEE.

关键词： distributed computing Edge computing Optimization

来源：评论

学校读者我要写书评

暂无评论

SEECSSim: A toolkit for parallel and distributed simulations for mobile devices

引用

JOURNAL OF SIMULATION 2021年第3期15卷 235-260页

作者： Maqbool, Fahad Malik, Asad Waqar Naqvi, Syed Meesam Raza Ahmed, Nadeem D'Angelo, Gabriele Mahmood, Imran Natl Univ Sci & Technol Sch Elect Engn & Comp Sci Islamabad Pakistan Univ Malaya Fac Comp Sci & Informat Technol Dept Informat Syst Kuala Lumpur Malaysia Univ Bologna Dept Comp Sci & Engn DISI Bologna Italy

Increase of the computing power on handheld devices has made these devices attractive for many applications including simulation systems. There are Parallel Discrete-Event Simulation (PDES) frameworks exists but most of these are designed for cluster systems, not suitable for battery operated devices. In this paper, the well- known Rensselaer's Optimistic Simulation System (ROSS) framework has been instrumented for a detailed analysis in terms of CPU usage, memory consumption, energy and power requirements. This profiling helps in selecting the appropriate synchronisation algorithm for running the PDES frameworks on the mobile devices. Further, we propose a simulation framework (referred as SEECSSim1) that is specifically designed for handheld devices. It is the first one designed keeping in mind the characteristics and the constraints of mobile devices. SEECSSim includes the support for a number of state-of-the-art synchronisation protocols. The performance of SEECSSim has been studied using a PHOLD simulation model for different synchronisation algorithms.

关键词： distributed computing mobile computing performance analysis simulations

来源：评论

学校读者我要写书评

暂无评论

Influencing Factors in the Scalability of distributed Stream Processing Jobs

引用

IEEE ACCESS 2021年 9卷 109413-109431页

作者： Van Dongen, Giselle van den Poel, Dirk Univ Ghent Dept MIO Data Analyt B-9000 Ghent Belgium

More and more use cases require fast, accurate, and reliable processing of large volumes of data. To do this, a distributed stream processing framework is needed which can distribute the load over several machines. In this work, we study and benchmark the scalability of stream processing jobs in four popular frameworks: Flink, Kafka Streams, Spark Streaming, and Structured Streaming. Besides that, we determine the factors that influence the performance and efficiency of scaling processing jobs with distinct characteristics. We evaluate horizontal, as well as vertical scalability. Our results show how the scaling efficiency is impacted by many factors including the initial cluster layout and direction of scaling, the pipeline design, the framework design, resource allocation, and data characteristics. Finally, we give some recommendations on how practitioners should undertake to scale their clusters.

关键词： Scalability Throughput Pipelines Benchmark testing Sparks Measurement Storms Apache spark structured streaming apache flink apache kafka kafka streams distributed computing stream processing frameworks scalability benchmarking big data

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：