检索结果-内蒙古大学图书馆

FPDeep: Scalable Acceleration of CNN Training on Deeply-Pipelined FPGA Clusters

IEEE TRANSACTIONS ON COMPUTERS 2020年第8期69卷 1143-1158页

作者： Wang, Tianqi Geng, Tong Li, Ang Jin, Xi Herbordt, Martin Univ Sci & Technol China Dept Phys Hefei 230026 Peoples R China Boston Univ Dept Elect & Comp Engn Boston MA 02215 USA Pacific Northwest Natl Lab PCSD Richland WA 99354 USA

Deep convolutional Neural Networks (CNNs) have revolutionized numerous applications, but the demand for ever more performance remains unabated. Scaling CNN computations to larger clusters is generally done by distributing tasks in batch mode using methods such as distributed synchronous SGD. Among the issues with this approach is that, to make the distributed cluster work with high utilization, the workload distributed to each node must be large;this implies nontrivial growth in the SGD mini-batch size. In this article we propose a framework, called FPDeep, which uses a hybrid of model and layer parallelism to configure distributed reconfigurable clusters to train CNNs. This approach has numerous benefits. First, the design does not suffer from performance loss due to batch size growth. Second, work and storage are balanced among nodes through novel workload and weight partitioning schemes. Part of the mechanism is the surprising finding that it is preferable to store excess weights in neighboring devices rather than in local off-chip memory. Third, the entire system is a fine-grained pipeline. This leads to high parallelism and utilization and also minimizes the time that features need to be cached while waiting for back-propagation. As a result, storage demand is reduced to the point where only on-chip memory is used for the convolution layers. And fourth, we find that the simplest topology, a 1D array, is preferred for interconnecting the FPGAs thus enabling widespread applicability. We evaluate FPDeep with the Alexnet, VGG-16, and VGG-19 benchmarks. Results show that FPDeep has good scalability to a large number of FPGAs, with the limiting factor being the FPGA-to-FPGA bandwidth. But with 250 Gb/s bidirectional bandwidth per FPGA, which is easily supported by current generation FPGAs, FPDeep performance shows linearity up to 100 FPGAs. Energy efficiency is evaluated with respect to GOPs/J. FPDeep provides, on average, 6.4x higher energy efficiency than comparabl

关键词： Training Field programmable gate arrays parallel processing Engines Pipelines Convolution Bandwidth Convolution neural network training reconfigurable computing parallel and distributed systems

来源：评论

学校读者我要写书评

暂无评论

parallel and distributed Kmeans to Identify the Translation Initiation Site of Proteins

Parallel and Distributed Kmeans to Identify the Translation ...

引用

IEEE International Conference on systems, Man, and Cybernetics

作者： Laerte M. Rodrigues Luis E. Zarate Cristiane N. Nobre Henrique C. Freitas Department of Computer Science Pontificia Universidade Catolica de Minas Gerais Department of Computer Science Universidade Federal de Sao Joao del-Rei

ISBN: (纸本)9781467317139

Prediction of the translation initiation site is of vital importance in bioinformatics since through this process it is possible to understand the organic formation and metabolic behavior of living organisms. Sequential algorithms are not always a viable solution due to the fact that mRNA databases are normally very large, resulting in long processing times. Applying parallel and distributed computing resources to such databases could help reduce this time. The objective of this article is to present a class balancing solution for the translation initiation site process using parallel and distributed computing resources in a hybrid model. The results reveal a speedup of up to 23 times compared to sequential methods and performance rates for accuracy, precision, sensitivity, specificity and adjusted accuracy of 91.15%, 39.83%, 89.11%, 88.93% and 89.02%, respectively, for the Homo sapiens database. For the Drosophila melanogaster database, the speedup was 18.33 times and accuracy, precision, sensitivity, specificity and adjusted accuracy were 95.22%, 43.01%, 90.83%, 90.47% and 90.64%, respectively. Both sets of results are considered important. Thus, the solution presented in this article demonstrated itself viable for the problem in question.

关键词： parallel and distributed systems Bioinformatics Clustering Translation Initiation parallel Lines Database Speeding Sensitivity and Specificity Proteins distributed energy resources Hybrid modes distributed computing Bioinformatics Anatomic Site Performance appraisal

来源：评论

学校读者我要写书评

暂无评论

Decentralized electric power grid architecture and management for reliable and cost-effective supply (DEZENT)

引用

WIRTSCHAFTSINFORMATIK 2007年第5期49卷 361-369页

作者： Wedde, Horst F. Lehnhoff, Sebastian Handschin, Edmund Krause, Olav Univ Dortmund Fachbereich Informat D-44221 Dortmund Germany Univ Dortmund Fak Elektrotech D-44221 Dortmund Germany

Given the sharply rising costs for traditional, energy (based on coal, fossil oil or gas etc). over the past few years renewable energy sources such as wind, sun, water, or, seed oil have favorably come into the picture as economically desirable and ecologically clean alternatives. The corresponding power facilities (e.g. wind power stations, solar panels, block beat & power-plants) are widely distributed and are quite heterogeneous regarding their productivity (capacity and reliability) yet at the same time transport paths and cost are at a minimal level. (In the near future it will be possible to cover all needs from such sources). For guaranteeing both the adaptive integration of the diverse facilities and a balanced level of supply under (locally) highly unpredictable energy production we present a bottom-up power grid management architecture. A key novelty of our approach is a completely decentralized management for negotiating the available power supply and needs. This is realized through a safety-critical, real-time multi-agent system where bids and offers are negotiated on the (electrically) shortest time basis of 0.5 sec. While the system is secure against malicious attacks it exhibits a high amount of fault tolerance where the latter corresponds well to the exceptional production safety of the widely distributed facilities. Beyond the high supply reliability under the decentralized management of distributed facilities we demonstrate that both operation costs and consumer prices could be assumed lower than under a centralized management and architecture.

关键词： renewable energy parallel and distributed systems SmartGrids multiagent systems autonomous software agents

来源：评论

学校读者我要写书评

暂无评论

Best of SBAC-PAD 2012

引用

parallel COMPUTING 2014年第9期40卷 512-513页

作者： Schnorr, Lucas Mello Alexandre Navaux, Philippe Olivier Univ Fed Rio Grande do Sul Inst Informat BR-91501970 Porto Alegre RS Brazil

This special issue presents new trends in computer architecture and in parallel and distributed systems. It is based on the best papers of the 24th International Symposium on Computer Architecture and High Performance Computing, which was held in New York, NY, USA on October 24-26, 2012 in the Columbia University. The authors were invited to provide extended versions of the papers presented in the conference, taking into account suggestions by the double-blinded peer review process and comments gathered during the conference.

关键词： Computer architecture parallel and distributed systems High performance computing

来源：评论

学校读者我要写书评

暂无评论

An empirical study of hierarchical division for mesh-structured networks

引用

JOURNAL OF COMPUTATIONAL METHODS IN SCIENCES AND ENGINEERING 2006年第5-6期6卷 S413-S426页

作者： Wang, Dajin Montclair State Univ Dept Comp Sci Montclair NJ 07043 USA

A parallel/distributed system consists of a collection of processes, which are distributed over a network of processors, and work in a cooperative manner to fulfill various tasks. A hierarchical approach is to group and organize the distributed processes into a logical hierarchy of multiple levels to achieve better system performance. It has been proposed as an effective way to solve various problems in distributed computing, such as distributed monitoring, resource scheduling, and network routing. In [21], we studied hierarchical configuration for mesh and hypercube networks to the end of achieving better system performance. In particular, we proposed theoretically optimal hierarchy for mesh and hypercube, so that the total traffic flow over the network is minimized. In this paper, we present the experimental results to establish the practical relevance of mesh hierarchy proposed in [21]. We simulated situations for multi-level division, real network loading scenarios, random data aggregation rates, and different division sizes other than derived in [21]. The simulation results not only show that the analytically obtained hierarchy works well for many realistic settings, but also offer some useful insights into the proposed hierarchy scheme.

关键词： Hierarchical architecture hierarchy interconnection networks mesh parallel and distributed systems simulation

来源：评论

学校读者我要写书评

暂无评论

New Model and Genetic Algorithm for Multi-Installment Divisible-Load Scheduling

New Model and Genetic Algorithm for Multi-Installment Divisi...

引用

IEEE Congress on Evolutionary Computation (CEC)

作者： Wang, Xiaoli Wang, Yuping Wei, Zhen Wei, Jingxuan Xidian Univ Sch Comp Sci & Technol Xian Peoples R China

ISBN: (纸本)9781479974924

The era of big data computing is coming. As scientific applications become more data intensive, finding an efficient scheduling strategy for massive computing in parallel and distributed systems has drawn increasingly attention. Most existing studies considered single-installment scheduling models, but very few literature involved multi-installment scheduling, especially in heterogeneous parallel and distributed systems. In this paper, we proposed a new model for periodic multi-installment divisible-load scheduling in which the make-span of the workload is minimized, and a genetic algorithm was designed to solve this model. Finally, experimental results show the effectiveness and efficiency of the proposed algorithm.

关键词： divisible load multi-installment scheduling global optimization genetic algorithm parallel and distributed systems

来源：评论

学校读者我要写书评

暂无评论

Assignment Algorithm for Energy Minimization on parallel Machines

Assignment Algorithm for Energy Minimization on Parallel Mac...

引用

38th International Conference on parallel Processing

作者： Kang, Jaeyeon Ranka, Sanjay Univ Florida Dept Comp & Informat Sci & Engn Gainesville FL 32611 USA

ISBN: (纸本)9781424449231

Energy consumption is a critical issue in parallel and distributed systems. Energy-Efficient Scheduling of Directed Acyclic Graph (DAG) based workflows on Dynamic Voltage Scaling (DVS) enabled systems consists of two phases: assignment of tasks and slack allocation. Most current research on scheduling for energy minimization of DAGs tries to minimize energy by effective slack allocation for a given assignment. The assignment itself does not take energy profiles of tasks into account. In this paper, we show that incorporating DVS based energy profiles of tasks during the assignment phase can lead to significantly lower overall energy requirements while requiring lower computational time.

关键词： parallel and distributed systems energy-efficient scheduling dynamic voltage scaling

来源：评论

学校读者我要写书评

暂无评论

On Signal Reconstruction Algorithms and Speedup Opportunities 34

On Signal Reconstruction Algorithms and Speedup Opportunitie...

引用

34th National Radio Science Conference (NRSC)

作者： Helal, Manal E. Ghouz, Hussein Hamed Arab Acad Sci Technol & Maritime Transport Engn & Technol Coll Cairo Egypt Brunel Univ London Inst Mat & Mfg London England Qassim Univ Coll Engn Al Malida Buraydah Saudi Arabia

ISBN: (纸本)9781509046096

Compressive Sensing (CS) reduces sampling data at the cost of increased signal reconstruction time. This problem has been addressed by a different set of algorithmic approaches under various realistic assumptions. Graphical Processing Units (GPU), Multicore architectures, distributed cloud computing are among technologies used to speed up data intensive and compute intensive algorithms. In this study, state of the art sample reconstruction algorithms (SRA) have been presented and surveyed. Opportunities to speed up their performances using parallel & distributed computing platforms have been also investigated and summarized. Finally, the study concludes with a detailed list of project ideas for further developments of these algorithms.

关键词： compressive sensing reconstruction algorithms parallel and distributed systems

来源：评论

学校读者我要写书评

暂无评论

A new approach to object oriented simulation of concurrent systems

A new approach to object oriented simulation of concurrent s...

引用

Conference on Enabling Technology for Simulation Science

作者： Pears, AN Singh, S Dillon, TS La Trobe University Bundoora Melbourne Vic. 3083 Australia

ISBN: (纸本)0819424986

High speed simulation of concurrent systems requires distributed processing if meaningful results are to be obtained for large systems in a reasonable timeframe. One of the most common methods used for such simulation is parallel Discrete Event Simulation (PDES). A range of PDES simulation kernels have been developed and much research has been devoted to optimistic execution strategies such as TimeWarp. Unfortunately in all this effort some fundamental aspects of object oriented modelling for simulation have received scant attention. In particular the ability of simulation kernels to act on truly generic simulation objects. In this context we define a truly generic object to be one which totally defines its responeses to external stimuli, but which has no concept of its place in the interconnected web of objects that comprise the total simulation environment. To address this problem, we propose a new modelling approach based on interacting objects, and an associated simulation kernel architecture. This paper describes the architecture and features of our simulation kernel in detail, and demonstrates, using a small example, the benefits of adopting such a modelling approach. The major specific benefits include true object genericity, enhanced scope for object re-use, and enhanced opportunities to use polymorphism.

关键词： object-oriented simulator generic distributed kernel parallel and distributed systems

来源：评论

学校读者我要写书评

暂无评论

What computer architecture can learn from computational intelligence - and vice versa

What computer architecture can learn from computational inte...

引用

23rd Euromicro Conference on New Frontiers of Information Technology

作者： Moore, R Klauer, B Waldschmidt, K J. W. Goethe-University Technische Informatik 60054 Frankfurt P.O. Box 111932 Germany

ISBN: (纸本)0818681306

This paper considers whether the seemingly disparate fields of Computational Intelligence (CI) and computer architecture can profit from each others' principles, results and experience. In the process, we identify important common issues, such as parallelism, distribution of data and control, granularity and regularity. We present two novel computer architectures which have profited from principles found in CI, and identify two constraints on CI to eliminate the hidden influence of the von Neumann model of computation.

关键词： Memory systems and management Multithreaded Architecture Neuro-Computing parallel and distributed systems parallelizing compilers

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：