基于网格的神经计算平台资源分配的设计与实现
|
摘要
神经网络的研究内容相当广泛，与多种科学领域的发展密切相关，目前在语音识别，字符识别，影像技术，图案识别和分类，信号过程，过程控制和优化等领域得到许多应用。正是因为越来越多的应用需要神经网络的支持，我们着手开发一个神经计算平台，该平台实现多种神经网络模型，例如BP模型、Hopfield模型、Boltzmann机模型、ART模型、BAM模型、遗传算法等等。而且能够根据用户的需要将多种模型进行结合计算。但是因为神经计算处理的数据比较庞大，所以为了提高运算速度，我们引进了网格计算(Grid Computing)的结构思想，架构一个分布式系统。
本文提出的是平台中资源分配系统的设计和实现。
系统用Globus项目中的信息服务——元计算目录服务MDS来发现资源。它主要是一种基于网格计算环境的信息服务框架，面向网格计算环境中资源数目巨大、地理上分布动态性的特点。MDS的内容主要包括资源(服务)发现、资源(服务)描述和资源(服务)监视与更新。通过MDS，本系统可以实时获取资源信息，满足自己的需求。
而且与传统的资源分配方法不同的是，本系统中采用的数据模型称为Classified Advertisement。这个模型非常灵活而且容易扩展。它的一个特点是把资源的查询封装到数据模型中，以属性的方式来表达。其中的Constrain属性就是机器对任务，或者是任务对机器的限制要求。而任务对服务质量的定义则在Rank属性中得到体现。这样，不同的用户可以根据自己的实际情况定义不同的Qos标准，更具合理性。Classad的另一个特点就是它也允许资源提供者自己来定义使用策略，对于不符合资源本身使用限制的那些任务，资源可以拒绝执行。这对于资源分布式所有的环境来说，赋予了机器选择的权利，使整个系统更具灵活性。
文章分五个部分，第一部分介绍神经网络和网格技术的发展，提出设计神经计算平台的意义；第二部分介绍了资源分配系统的设计和实现，包括主要模块，映射算法和如何激活特定任务；文章第三部分介绍了Globus项目中的MDS的使用；第四部分详细描述Classified Advertisement的数据结构。最后的第五和第六部分是运行实例与文章的总结。
Neural networks have inspired many scientists to propose them as a solution for various problems because they have such good properties as parallel functioning, relatively easy implementation of complicated tasks, distributed information storage and learning abilities. Ideally these networks could accomplish an arbitrary task with correct topology and sufficient training. Many models, such as Back Propagation (BP), Hopfield, ART, have been developed and sometimes several models have to be combined to accomplish a task. To relieve the burden of implementing those models from scratch, we developed a neural computation platform (NCP) containing those facilities. The training of a particular neural network involves huge amount of data. To improve the speed of computation, we used the idea of Grid Computing to construct a distributed system.
The purpose of this paper is to present an autonomous resource allocation method used in NCP.
NCP uses Metacomputing Directory Service (MDS) to discover the resources. MDS provides information services in Globus project. It adapts to the environment which has tremendous resources and services and the resources are distributed owned. MDS is designed to provide a standard mechanism for publishing and discovering resource status and configuration information. It provides a uniform, flexible interface to data collected by lower-level information providers. It has a decentralized structure that allows it to scale, and it can handle static or dynamic data. NCP gets the resource information all the time by the MDS.
Traditional resource allocation method requires a unified system and the allocation is carried on by fixed rules. Usually the shortest task execution time is prefered. But in the Grid environment, user actions and resource conditions are very complex, so only pursuiting computation speed can not satisfy all conditions. Instead, quality of service (QoS) is an important aspect to tasks. To accommodate such requirement, a semi-structure data model called Classified Advertisement (Classad) is adopted by NCP. Classad is flexible and extensive, which encapsulates resource queries into the data model. A Classad may contain the following items: attribute, constraint, rank. Constraint exists between computing nodes and tasks. They can both set limits to the other. Rank embodies the definition of QoS from tasks so that users can define different QoS rules under different circumstances. The most outsanding
characteristic of Classad is that it allows computing nodes to define their own policies. Any tasks conflict with policies will be rejected by those computing nodes.
The thesis has five chapters. The first chapter, we introduce the developement of the Neural Network and Grid technology. The second chapter introduces the design and implement of the resource allocation system including the main moduls, mapping alorgithm and the activation of the program. The third chapter introduces how to use the MDS. The fouth chapter introduces the data structure of Classified Advertisement. The fouth chapter demonstrates the example and in the last chapter, we summarize the whole thesis.
本文提出的是平台中资源分配系统的设计和实现。
系统用Globus项目中的信息服务——元计算目录服务MDS来发现资源。它主要是一种基于网格计算环境的信息服务框架，面向网格计算环境中资源数目巨大、地理上分布动态性的特点。MDS的内容主要包括资源(服务)发现、资源(服务)描述和资源(服务)监视与更新。通过MDS，本系统可以实时获取资源信息，满足自己的需求。
而且与传统的资源分配方法不同的是，本系统中采用的数据模型称为Classified Advertisement。这个模型非常灵活而且容易扩展。它的一个特点是把资源的查询封装到数据模型中，以属性的方式来表达。其中的Constrain属性就是机器对任务，或者是任务对机器的限制要求。而任务对服务质量的定义则在Rank属性中得到体现。这样，不同的用户可以根据自己的实际情况定义不同的Qos标准，更具合理性。Classad的另一个特点就是它也允许资源提供者自己来定义使用策略，对于不符合资源本身使用限制的那些任务，资源可以拒绝执行。这对于资源分布式所有的环境来说，赋予了机器选择的权利，使整个系统更具灵活性。
文章分五个部分，第一部分介绍神经网络和网格技术的发展，提出设计神经计算平台的意义；第二部分介绍了资源分配系统的设计和实现，包括主要模块，映射算法和如何激活特定任务；文章第三部分介绍了Globus项目中的MDS的使用；第四部分详细描述Classified Advertisement的数据结构。最后的第五和第六部分是运行实例与文章的总结。
Neural networks have inspired many scientists to propose them as a solution for various problems because they have such good properties as parallel functioning, relatively easy implementation of complicated tasks, distributed information storage and learning abilities. Ideally these networks could accomplish an arbitrary task with correct topology and sufficient training. Many models, such as Back Propagation (BP), Hopfield, ART, have been developed and sometimes several models have to be combined to accomplish a task. To relieve the burden of implementing those models from scratch, we developed a neural computation platform (NCP) containing those facilities. The training of a particular neural network involves huge amount of data. To improve the speed of computation, we used the idea of Grid Computing to construct a distributed system.
The purpose of this paper is to present an autonomous resource allocation method used in NCP.
NCP uses Metacomputing Directory Service (MDS) to discover the resources. MDS provides information services in Globus project. It adapts to the environment which has tremendous resources and services and the resources are distributed owned. MDS is designed to provide a standard mechanism for publishing and discovering resource status and configuration information. It provides a uniform, flexible interface to data collected by lower-level information providers. It has a decentralized structure that allows it to scale, and it can handle static or dynamic data. NCP gets the resource information all the time by the MDS.
Traditional resource allocation method requires a unified system and the allocation is carried on by fixed rules. Usually the shortest task execution time is prefered. But in the Grid environment, user actions and resource conditions are very complex, so only pursuiting computation speed can not satisfy all conditions. Instead, quality of service (QoS) is an important aspect to tasks. To accommodate such requirement, a semi-structure data model called Classified Advertisement (Classad) is adopted by NCP. Classad is flexible and extensive, which encapsulates resource queries into the data model. A Classad may contain the following items: attribute, constraint, rank. Constraint exists between computing nodes and tasks. They can both set limits to the other. Rank embodies the definition of QoS from tasks so that users can define different QoS rules under different circumstances. The most outsanding
characteristic of Classad is that it allows computing nodes to define their own policies. Any tasks conflict with policies will be rejected by those computing nodes.
The thesis has five chapters. The first chapter, we introduce the developement of the Neural Network and Grid technology. The second chapter introduces the design and implement of the resource allocation system including the main moduls, mapping alorgithm and the activation of the program. The third chapter introduces how to use the MDS. The fouth chapter introduces the data structure of Classified Advertisement. The fouth chapter demonstrates the example and in the last chapter, we summarize the whole thesis.
引文
[1]. I.Foster and C.Kesselman, The Grid: Blueprint for A New Computing Infrastructure. Morgan Kaufmann, 1999. 159-180.
[2]. The Globus Project http://www.globus.org.
[3]. Czajkowski,K., Fitzgerald,S., Foster, I. and Kesselman,C., Grid Information Services for Distributed Resource Sharing. In 10~(th) IEEE International Symposium on High Performance Distributed Computing, (2001),IEEE Press, 181-184
[4]. Fitzgerald, S., Foster, I., Kesselman,C., Laszewski,G.v., Smith,W. and Tuecke, S. A Directory Service for configuring High-performance Distributed Computations. In Proc.6~(th) IEEE Symp.on High Performance Distributed Computing, 1997, 365-375.
[5]. Raman, R., Livny, M. and Solomon, M., Matchmaking: Distributed Resource Management for High Throughput Computing. In IEEE international Symposium on High Performance Distributed Computing, (1998), IEEE Press, 140-147.
[6]. R. Raman, "Matchmaking Frameworks for Distributed Resource Management". PhD thesis, Department of Computer Science, University of Wisconsin - Madison, 2001.
[7]. Muthucumaru Maheswaran, Howard Jay Siegel, A Dynamic Matching and Scheduling Algorithm for Heterogeneous Computing Systems. In: Proceedings of the 7~(th) IEEE Heterogeneous Computing Workshop (HCW' 98), IEEE Computer Society Press, 57-69.
[8]. Howard Jay Siegel, Shoukat Ali, Techniques for Mapping Tasks to Machines in Heterogeneous Computing Systems, Euromicro Journal of Systems Architecture, Special Issue on Heterogeneous Distributed and Parallel Architecture: Hardware, Software an Design Tools, Vol.46, No.8, 2000, 627-639.
[9]． 《网格计算》，都志辉，陈渝，刘鹏，清华大学出版社。ISBN 7-302-05939-X／TP．3531，2002．10
[10]. Cray, R.Document number in-2153 2/97. Cray Research, 1997
[11]. Henderson,R. and Tweten, D. Portable Batch System: External reference specification. Ames Research Center, 1996
[12]. Raman, R. ClassAds Programming Tutorial(C++).2000
[13]. I.Foster and C. Kesselman. Globus: A metacomputing infrastructure toolkit. Intl.Journal of Supercomputing Applications, 11(2): 115-128,1997
[14]. MDS 2.2 User's Guide
[15]. P.Dinda and B.Plale. A unified relational approach to grid information services. Grid Forum Working Draft GWD-GIS-012-1, February 2001.
http://www.gridforum.org
[16]. I.Foster, C. Kesselman, and S. Tuecke. The anatomy of the Grid: Enabling scalable virtual organizations. Intl. Journal of Supercomputing Applications, (to appear) 2001.
[17]. S.Fitzgerald, I. Foster, C. Kesselman, G. von Laszewski, W. Smith, and S. Tuecke. A directory service for configuring high-performance distributed computations. In Proc. 6~(th) IEEE Symp. On High Performance Distributed Computing, pages 365-375, 1997
[18]. M. Smith, G.Good, T. Howes, and R. Weltman. Persistent search: A simple LDAP change notification mechanism. Working document, November 2000.
[19]. E Stelling, I. Foster, C. Kesselman, C. Lee, and G. vonLaszewski. A fault detection service for wide area distributed computations. In Proc. 7~(th) IEEE Symp. On High Performance Distributed Computing, pages 268-278, 1998
[20]. B. Tierney, B. Crowley, D. Gunter, M. Holding, J. Lee, and M. Thompson. A monitoring sensor management system for grid environments. In Proc. 9~(th) IEEE Symp. On High Performance Distributed Computing, pages 97-104, 2000
[21]. Foster, I. and Kesselman, C., "Globus: A Toolkit-Based Grid Architecture".
[22]. Abramson, D., Giddy, J., and Kotler, L., "High Performance Parametric Modeling with Nimrod/G: Killer Application for the Global Grid?", IPDPS'2000, IEEE Press,2000
[23]. Berman, F., "High-Performance Schedulers", Foster, I. and Kesselman, C. eds
[24]. Czajkowski, K., Fitzgerald, S., Foster, I., and Kesselman, C., "Grid Informaiton Services for Distributed Resource Sharing", 2001
[25]. Epema, D.H.J., Livny, M., Danzig, R.v., Evers, X., and Prutne, J., "A Worldwide Flock of Condors: Load Sharing among Workstation Clusters", Future Generation Computer Systems, 12, 1996
[26]. Foster I. and Kesselman, C. Globus: A Toolkit-Based Grid Architecture. In Foster, I. and Kesselman, C. eds. The Grid." Blueprint for a New Computing Infrastructure, Morgan Kaufmann 1999, 259-278.
[27]. Foster I., Kesselman, C., Lee, C., Lindell, R., Nahrstedt, K. and Roy, A., A Distributed Resource Management Architecture that Supports Advance Reservations and Co-Allocation. In Proc. International Workshop on Quality of Service, (1999), 27-36
[28]. Gannon, D., Bramley, R., Fox, G., Smallen, S., Rossi, A., Ananthakrisnan, R., Bertrand, F., Chiu, K., Farrellee, M., Govindaraju, M., Krishnan, S., Ramakrishnan, L., Simmhan, Y., Slominski, A., Ma, Y., Olariu, C. and Rey-Cenvaz, N., Programming the Grid: Distributed Software Components, P2P, and Grid Web
Services for Scientific Applications. In Grid 2001, (2001)
[29]. Czajkowski, K., Foster, I., Karonis, N., Kesselman, C., Martin, S., Smith, W. and Tuecke, S. A Resource Management Architecture for Metacomputing Systems. In 4th Workshop on Job Scheduling Strategies for Parallel Processing, Springer-Verlag, 1998, 62-82.
[30]. Bramley, R., Gannon, D., Stuckey, T., Villacis, J., Balasubramanian, J., E. Akman, Breg, F., Diwan, S. and Govindaraju, M. Component Architectures for Distributed Scientific Problem Solving. IEEE Computational Science and Engineering, 5(2). 50-63. 1998.
[2]. The Globus Project http://www.globus.org.
[3]. Czajkowski,K., Fitzgerald,S., Foster, I. and Kesselman,C., Grid Information Services for Distributed Resource Sharing. In 10~(th) IEEE International Symposium on High Performance Distributed Computing, (2001),IEEE Press, 181-184
[4]. Fitzgerald, S., Foster, I., Kesselman,C., Laszewski,G.v., Smith,W. and Tuecke, S. A Directory Service for configuring High-performance Distributed Computations. In Proc.6~(th) IEEE Symp.on High Performance Distributed Computing, 1997, 365-375.
[5]. Raman, R., Livny, M. and Solomon, M., Matchmaking: Distributed Resource Management for High Throughput Computing. In IEEE international Symposium on High Performance Distributed Computing, (1998), IEEE Press, 140-147.
[6]. R. Raman, "Matchmaking Frameworks for Distributed Resource Management". PhD thesis, Department of Computer Science, University of Wisconsin - Madison, 2001.
[7]. Muthucumaru Maheswaran, Howard Jay Siegel, A Dynamic Matching and Scheduling Algorithm for Heterogeneous Computing Systems. In: Proceedings of the 7~(th) IEEE Heterogeneous Computing Workshop (HCW' 98), IEEE Computer Society Press, 57-69.
[8]. Howard Jay Siegel, Shoukat Ali, Techniques for Mapping Tasks to Machines in Heterogeneous Computing Systems, Euromicro Journal of Systems Architecture, Special Issue on Heterogeneous Distributed and Parallel Architecture: Hardware, Software an Design Tools, Vol.46, No.8, 2000, 627-639.
[9]． 《网格计算》，都志辉，陈渝，刘鹏，清华大学出版社。ISBN 7-302-05939-X／TP．3531，2002．10
[10]. Cray, R.Document number in-2153 2/97. Cray Research, 1997
[11]. Henderson,R. and Tweten, D. Portable Batch System: External reference specification. Ames Research Center, 1996
[12]. Raman, R. ClassAds Programming Tutorial(C++).2000
[13]. I.Foster and C. Kesselman. Globus: A metacomputing infrastructure toolkit. Intl.Journal of Supercomputing Applications, 11(2): 115-128,1997
[14]. MDS 2.2 User's Guide
[15]. P.Dinda and B.Plale. A unified relational approach to grid information services. Grid Forum Working Draft GWD-GIS-012-1, February 2001.
http://www.gridforum.org
[16]. I.Foster, C. Kesselman, and S. Tuecke. The anatomy of the Grid: Enabling scalable virtual organizations. Intl. Journal of Supercomputing Applications, (to appear) 2001.
[17]. S.Fitzgerald, I. Foster, C. Kesselman, G. von Laszewski, W. Smith, and S. Tuecke. A directory service for configuring high-performance distributed computations. In Proc. 6~(th) IEEE Symp. On High Performance Distributed Computing, pages 365-375, 1997
[18]. M. Smith, G.Good, T. Howes, and R. Weltman. Persistent search: A simple LDAP change notification mechanism. Working document, November 2000.
[19]. E Stelling, I. Foster, C. Kesselman, C. Lee, and G. vonLaszewski. A fault detection service for wide area distributed computations. In Proc. 7~(th) IEEE Symp. On High Performance Distributed Computing, pages 268-278, 1998
[20]. B. Tierney, B. Crowley, D. Gunter, M. Holding, J. Lee, and M. Thompson. A monitoring sensor management system for grid environments. In Proc. 9~(th) IEEE Symp. On High Performance Distributed Computing, pages 97-104, 2000
[21]. Foster, I. and Kesselman, C., "Globus: A Toolkit-Based Grid Architecture".
[22]. Abramson, D., Giddy, J., and Kotler, L., "High Performance Parametric Modeling with Nimrod/G: Killer Application for the Global Grid?", IPDPS'2000, IEEE Press,2000
[23]. Berman, F., "High-Performance Schedulers", Foster, I. and Kesselman, C. eds
[24]. Czajkowski, K., Fitzgerald, S., Foster, I., and Kesselman, C., "Grid Informaiton Services for Distributed Resource Sharing", 2001
[25]. Epema, D.H.J., Livny, M., Danzig, R.v., Evers, X., and Prutne, J., "A Worldwide Flock of Condors: Load Sharing among Workstation Clusters", Future Generation Computer Systems, 12, 1996
[26]. Foster I. and Kesselman, C. Globus: A Toolkit-Based Grid Architecture. In Foster, I. and Kesselman, C. eds. The Grid." Blueprint for a New Computing Infrastructure, Morgan Kaufmann 1999, 259-278.
[27]. Foster I., Kesselman, C., Lee, C., Lindell, R., Nahrstedt, K. and Roy, A., A Distributed Resource Management Architecture that Supports Advance Reservations and Co-Allocation. In Proc. International Workshop on Quality of Service, (1999), 27-36
[28]. Gannon, D., Bramley, R., Fox, G., Smallen, S., Rossi, A., Ananthakrisnan, R., Bertrand, F., Chiu, K., Farrellee, M., Govindaraju, M., Krishnan, S., Ramakrishnan, L., Simmhan, Y., Slominski, A., Ma, Y., Olariu, C. and Rey-Cenvaz, N., Programming the Grid: Distributed Software Components, P2P, and Grid Web
Services for Scientific Applications. In Grid 2001, (2001)
[29]. Czajkowski, K., Foster, I., Karonis, N., Kesselman, C., Martin, S., Smith, W. and Tuecke, S. A Resource Management Architecture for Metacomputing Systems. In 4th Workshop on Job Scheduling Strategies for Parallel Processing, Springer-Verlag, 1998, 62-82.
[30]. Bramley, R., Gannon, D., Stuckey, T., Villacis, J., Balasubramanian, J., E. Akman, Breg, F., Diwan, S. and Govindaraju, M. Component Architectures for Distributed Scientific Problem Solving. IEEE Computational Science and Engineering, 5(2). 50-63. 1998.
© 2004-2018 中国地质图书馆版权所有 京ICP备05064691号 京公网安备11010802017129号 地址：北京市海淀区学院路29号 邮编：100083 电话：办公室：(+86 10)66554848；文献借阅、咨询服务、科技查新：66554700 |