检索结果-内蒙古大学图书馆

A parallel and distributed stochastic gradient descent implementation using commodity clusters

JOURNAL OF BIG DATA 2019年第1期6卷 1页

作者： Kennedy, Robert K. L. Khoshgoftaar, Taghi M. Villanustre, Flavio Humphrey, Timothy Florida Atlantic Univ 777 Glades Rd Boca Raton FL 33431 USA LexisNexis Business Informat Solut 245 Peachtree Ctr Ave Atlanta GA 30303 USA

Deep Learning is an increasingly important subdomain of artificial intelligence, which benefits from training on Big Data. The size and complexity of the model combined with the size of the training dataset makes the training process very computationally and temporally expensive. Accelerating the training process of Deep Learning using cluster computers faces many challenges ranging from distributed optimizers to the large communication overhead specific to systems with off the shelf networking components. In this paper, we present a novel distributed and parallel implementation of stochastic gradient descent (SGD) on a distributed cluster of commodity computers. We use high-performance computing cluster (HPCC) systems as the underlying cluster environment for the implementation. We overview how the HPCC systems platform provides the environment for distributed and parallel Deep Learning, how it provides a facility to work with third party open source libraries such as TensorFlow, and detail our use of third-party libraries and HPCC functionality for implementation. We provide experimental results that validate our work and show that our implementation can scale with respect to both dataset size and the number of compute nodes in the cluster.

关键词： parallel stochastic gradient descent parallel and distributed processing Deep learning Big data Neural network Cluster computer HPCC systems

来源：评论

学校读者我要写书评

暂无评论

The PN-tree: A parallel and distributed multidimensional index

引用

distributed AND parallel DATABASES 2005年第2期17卷 111-133页

作者： Ali, MH Saad, AA Ismail, MA Purdue Univ Dept Comp Sci W Lafayette IN 47907 USA Univ Alexandria Fac Engn Dept Comp Sci & Automat Control Alexandria Egypt

Multidimensional indexing is concerned with the indexing of multi-attributed records, where queries can be applied. on some or all of the. attributes. Indexing multi-attributed records is referred to by the term multidimensional indexing because each record is viewed as a point in a multidimensional space with a number of dimensions that is equal to the number of attributes. The values of the point coordinates along each dimension are equivalent to the values of the corresponding attributes. In this paper, the PN-tree, a new index structure for multidimensional spaces, is presented. This index structure is an efficient structure for indexing multidimensional points and is parallel by nature. Moreover, the proposed index structure does not lose its efficiency if it is serially processed or if it is processed using a small number of processors. The PN-tree can take advantage of as many processors as the dimensionality of the space. The PN-tree makes use of B(+)-trees that have been developed and tested over years in many DBMSs. The PN-tree is compared to the Hybrid tree that is known for its superiority among various index structures. Experimental results show that parallel processing of the PN-tree reduces significantly the number of disk accesses involved in the search operation. Even in its serial case, the PN-tree outperforms the Hybrid tree for large database sizes.

关键词： multidimensional indexing multi-attributed records parallel and distributed processing

来源：评论

学校读者我要写书评

暂无评论

Inconstant Update of Reference Point Value for parallel and distributed MOEA/D

Inconstant Update of Reference Point Value for Parallel and ...

引用

IEEE Congress on Evolutionary Computation (IEEE CEC)

作者： Sato, Mikiko Sato, Yuji Midtlyng, Mads Miyakawa, Minami Tokai Univ Sch Informat & Telecommun Engn Tokyo Japan Hosei Univ Fac Comp & Informat Sci Tokyo Japan Hosei Univ Grad Sch Comp & Informat Sci Tokyo Japan Shinshu Univ Fac Engn Nagano Japan

ISBN: (纸本)9781728183923

This paper proposes an update method to determine the reference point and nadir point when speeding up multi-objective evolutionary optimization algorithm based on decomposition (MOEA/D) with parallel and distributed processing in a many-core environment. We have already proposed a migration method that uses a virtual overlapping zone between partitions. Compared with simply partitioning the objective space or applying the standard island migration model, this method is shown to be effective in improving the hypervolume (HV) value, increasing diversity, and reducing regions where non-dominated solutions do not exist between partitions (sparse areas). However, this proposed method could not completely deal with the sparse areas' problem in the early stage of solution searching. To resolve this problem, in addition to the virtual overlap method, we propose here a method that switches the update interval for determining the reference point and nadir point by taking into account the progress of solution searching. To evaluate our newly proposed method, we compare both convergence and diversity on single-core MOEA/D and parallel MOEA/D using a two-objective constrained knapsack problem, and show that our method is effective not only in reducing sparse areas in the early stage but also in improving diversity and increasing the HV value.

关键词： MOEA/D parallel and distributed processing reference point multi-objective evolutionary algorithms

来源：评论

学校读者我要写书评

暂无评论

WSN-ANN: parallel and distributed Neurocomputing with Wireless Sensor Networks

WSN-ANN: Parallel and Distributed Neurocomputing with Wirele...

引用

International Joint Conference on Neural Networks (IJCNN)

作者： Serpen, Gursel Li, Jiakai Liu, Linqian Gao, Zhenning Univ Toledo Elect Engn & Comp Sci Dept Toledo OH 43606 USA

ISBN: (纸本)9781467361293;9781467361286

This paper proposes wireless sensor networks as a parallel and distributed computing platform for neurocomputing. The proposal entails leveraging the existing wireless sensor networks technology to serve as a hardware-software platform to implement and realize artificial neural network algorithms in fully parallel and distributed computation mode. The study describes the proposed parallel and distributed neurocomputing architecture, which is named as WSN-ANN, and its use as a hardware platform on a case study. A Hopfield neural network, which is configured to solve the minimum weakly connected dominating set problem, is embedded into a wireless sensor network. Simulation study results indicate that the proposed computing platform based on wireless sensor networks, WSN-ANN, is feasible and promising to serve as a parallel and distributed neurocomputer.

关键词： artificial neural network wireless sensor network parallel and distributed processing parallel hardware parallel and distributed computer architecture Hopfield neural network graph optimization

来源：评论

学校读者我要写书评

暂无评论

Adaptive PBI for Massively parallel MOEA/D in a distributed Memory Environment

Adaptive PBI for Massively Parallel MOEA/D in a Distributed ...

引用

IEEE Congress on Evolutionary Computation (CEC)

作者： Sato, Yuji Hirayama, Tomoya Ikami, Ryo Hosei Univ Fac Comp & Informat Sci Tokyo Japan

ISBN: (数字)9781665467087

ISBN: (纸本)9781665467087

This paper proposes an adaptive PBI for massively parallel MOEA/D in a distributed memory environment. Massively parallelization in a distributed memory environment effectively speeds up evolutionary multi-objective optimization algorithms for practical application problems. On the other hand, when MOEA/D is divided for parallelization by focusing on the reference vector in the objective function space, the T-neighbor is divided and the problem that the solution distribution becomes sparse near the boundary of the divided region arises. Here, we propose a method to improve the problem that the T-neighbor is divided and the solution distribution becomes sparse by adaptively controlling the penalty value in the PBI function according to the distance from the reference vector using a distribution function such as Laplace distribution. The effectiveness of the proposed method is shown by comparison with execution using a single CPU.

关键词： MOEA/D parallel and distributed processing distributed memory environment multi-objective evolutionary algorithms

来源：评论

学校读者我要写书评

暂无评论

Performance evaluation of RHiNET-2/NI: A network interface for distributed parallel computing systems

Performance evaluation of RHiNET-2/NI: A network interface f...

引用

3rd IEEE/ACM International Symposium on Cluster Computing and the Grid (CCGrid2003)

作者： Watanabe, K Otsuka, T Tsuchiya, JI Amano, H Harada, H Yamamoto, J Nishi, H Kudoh, T Keio Univ Dept Comp Sci Yokohama Kanagawa 2238522 Japan

ISBN: (纸本)0769519199

RHiNET-2/NI is a network interface for a parallel and distributed computing system with network connected PCs. The core of the network interface is an ASIC network controller chip Martini, which provides low-latency and large-bandwidth communication. Evaluation results show that it achieves almost full bandwidth of the 66MHz/64bit PCI bus, which is much larger than that of Myrinet-2000. The performance of a small prototype parallel system achieves almost linear speed up.

关键词： PC clusters high performance networks parallel and distributed processing performance evaluation SCore

来源：评论

学校读者我要写书评

暂无评论

parallel and distributed MOEA/D with Exclusively Evaluated Mating and Migration

Parallel and Distributed MOEA/D with Exclusively Evaluated M...

引用

IEEE Congress on Evolutionary Computation (CEC) as part of the IEEE World Congress on Computational Intelligence (IEEE WCCI)

作者： Sato, Yuji Sato, Mikiko Midtlyng, Mads Miyakawa, Minami Hosei Univ Fac Comp & Informat Sci Tokyo Japan Tokai Univ Sch Informat & Telecommun Engn Tokyo Japan Hosei Univ Grad Sch Comp & Informat Sci Tokyo Japan Sinshu Univ Dept Elect Informat Syst Engn Nagano Japan

ISBN: (纸本)9781728169293

This paper proposes a method for many-core-based large-scale parallel and distributed computation of MOEA/D, a decomposition-based evolutionary multi-objective optimization algorithm. Standard parallel MOEA/D on many-core environments provides fast execution time, but uniformity and diversity of the Pareto front may be lost. To avoid this problem, we propose a method of defining a virtual overlapping zone between partitions and selecting individuals for mating and migration by evaluating individual populations in this area using weight vectors of adjacent partitions. Using a two-objective constrained knapsack problem for evaluation, we compare the proposed method with standard single-core execution, no-migration parallel MOEA/D, and parallel MOEA/D with standard migration, and show that the proposed method is effective in improving diversity in solution searching while shortening execution time and increasing the accuracy of solution searching.

关键词： MOEA/D parallel and distributed processing many-core environment multi-objective evolutionary algorithms

来源：评论

学校读者我要写书评

暂无评论

Share-Frequent Sensor Patterns Mining from Wireless Sensor Network Data

引用

IEEE TRANSACTIONS ON parallel AND distributed SYSTEMS 2015年第12期26卷 3471-3484页

作者： Rashid, Md. Mamunur Gondal, Iqbal Kamruzzaman, Joarder Monash Univ Fac Informat Technol Clayton Vic 3800 Australia Federat Univ Mount Helen Vic Australia Monash Univ Clayton Vic 3800 Australia

Mining interesting knowledge from the huge amount of data gathered from WSNs is a challenge. Works reported in literature use support metric-based sensor association rules which employ the occurrence frequency of patterns as criteria. However, consideration of the binary frequency of a pattern is not a sufficient indicator for finding meaningful patterns because it only reflects the number of epochs which contain that pattern in the dataset. The share measure of sensorsets could discover useful knowledge about trigger values associated with a sensor. Here, we propose a new type of behavioral pattern called share-frequent sensor patterns (SFSPs) by considering the non-binary frequency values of sensors in epochs. SFSPs can find a correlation among a set of sensors and hence can improve the performance of WSNs in a resource management process. In this paper, a share-frequent sensor pattern tree (ShrFSP-tree) has been proposed to facilitate a pattern growth mining technique to discover SFSPs from WSN data. We also present a parallel and distributed method where the ShrFSP-tree is enhanced into PShrFSP-tree and its performance is investigated for both homogeneous and heterogeneous systems. Results show that our method is time and memory efficient in finding SFSPs than the existing most efficient algorithms.

关键词： parallel and distributed processing wireless sensor networks data mining knowledge discovery behavioral patterns share-frequent sensor patterns

来源：评论

学校读者我要写书评

暂无评论

Improvement of job completion time in data-intensive cloud computing applications

引用

JOURNAL OF CLOUD COMPUTING-ADVANCES SYSTEMS AND APPLICATIONS 2020年第1期9卷 1-20页

作者： Ibrahim, Ibrahim Adel Bassiouni, Mostafa Univ Cent Florida Dept Comp Engn Orlando FL 32816 USA Univ Technol Baghdad Dept Comp Engn Baghdad Iraq Univ Cent Florida Dept Comp Sci Orlando FL 32816 USA

Task stragglers in MapReduce jobs dramatically impede job execution of data-intensive computing in cloud data centers. This impedance is due to the uneven distribution of input data, heterogeneous data nodes, resource contention situations, and network configurations. Data skew of intermediate data in MapReduce job causes delay failures due to the violation of job completion time. Data-intensive computing frameworks, such as MapReduce or Hadoop YARN, employ HashPartitioner. This partitioner may cause intermediate data skew, which results in straggler reducers. In this paper, we strive to make Hadoop YARN more efficient in cloud environments. We present, a new partitioning scheme, called balanced data clusters partitioner (BDCP), to handle straggler Reduce tasks based on sampling of input data and feedback information about the current processing task. Our extensive experimental results show that BDCP can outperform the default Hadoop HashPartitioner and Range partitioner. BDCP can assist in straggler mitigation during reduce phase and minimize the job completion time in MapReduce jobs within data-intensive cloud computing.

关键词： Cloud computing MapReduce Data-intensive computing parallel and distributed processing Straggler reduce task Sampling

来源：评论

学校读者我要写书评

暂无评论

History-driven dynamic load balancing for recurring applications on networks of workstations

引用

JOURNAL OF SYSTEMS AND SOFTWARE 2000年第1期51卷 61-72页

作者： Bozyigit, M King Fahd Univ Petr & Minerals Dept Informat & Comp Sci Dhahran 31261 Saudi Arabia

The existing dynamic approaches to load balancing on parallel or distributed computer systems are primarily based on the current system load. Often, variations in the future exertable load are neglected or assumed as fixed. In this study. a new dynamic load balancing scheme (DYLOBA) is presented, where both the current system load and the load to be exerted by the application are equally important. The target system chosen is a general purpose network of workstations (NOW). The approach utilizes the past execution statistics of the applications. In this sense, information on the run time system load and resource requirement of the applications, averaged over past executions, is integrated. DYLOBA subsumes that most demanding applications are recurring. Thus, in the long run, it will be in possession of the execution histories of most applications. The system load balances both independent and parallel applications. Once an application joins DYLOBA it is permanently given a unique identification, to monitor it during future executions. DYLOBA was subjected to selective artificial parallel applications. The performance results under DYLOBA showed significant improvements over three other, rather conventional, task assignment schemes: random, fixed, and worst. (C) 2000 Elsevier Science Inc. All rights reserved.

关键词： dynamic load balancing recurring applications parallel and distributed processing networks of workstations

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：