Mapping of coarse-grained applications onto workstation-clusters.pdf

下载文档

0
0
约3.76万字
约 8页
2017-04-06 发布于江苏
举报
版权申诉
保障服务

Mapping of coarse-grained applications onto workstation-clusters.pdf

1、本文档共8页，可阅读全部内容。
2、有哪些信誉好的足球投注网站（book118）网站文档一经付费（服务费），不意味着购买了该文档的版权，仅供个人/单位学习、研究之用，不得用于商业用途，未经授权，严禁复制、发行、汇编、翻译或者网络传播等，侵权必究。
3、本站所有内容均由合作方或网友上传，本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺！文档内容仅供研究参考，付费前请自行鉴别。如您付费，意味着您自己接受本站规则且自行承担风险，本站不退款、不进行额外附加服务；查看《如何避免下载的几个坑》。如果您已付费下载过本站文档，您可以点击这里二次下载。
4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等，请点击“版权申诉”（推荐），也可以打举报电话：400-050-0827(电话支持时间：9:00-18:30)。

Mapping of coarse-grained applications onto workstation-clusters

Proceedings of the 5th EUROMICROWorkshop on Parallel and Distributed Processing, PDP 97, 1997, to appear. Mapping of Coarse-Grained Applications onto Workstation-Clusters Thomas Decker and Ralf DiekmannDepartment of Mathematics and Computer ScienceUniversity of Paderborn, D-33095 Paderborn, Germany http://www.uni-paderborn.de/cs/ag-monien.html e-mail: fdecker, diekg@uni-paderborn.de Abstract We present an environment for configuring and co- ordinating coarse grained parallel applications on work- station clusters. The environment named CoPA is based on PVM and allows an automatic distribution of functional modules as they occur in typical CAE-applications. By implementing link-based communication on top of PVM, CoPA is able to perform a “post-game” analysis of the communication load between different modules. Together with the computational load which is also determined au- tomatically by CoPA, all necessary information is avail- able to calculate an optimized mapping for the next run of the application. To optimize the distribution of modules onto workstations CoPA uses state-of-the-art partitioning heuristics as well as Simulated Annealing. Measurements show the large improvements in running time obtained by using optimization heuristics to determine the mapping. 1 Introduction A balanced distribution of data objects or programmodules is one of the most important premises for the efficient use of parallel and distributed systems. Using massively parallel systems with distributed memory to solve complex problems is a well researched area. Much work has been done on efficiently utilizing multi-processor parallel systems (MPPs). MPPs are usually scalable up to nearly any numbers of processors.The most popular pro- gramming model used within such systems is Single Pro- gram, Multiple Data (SPMD) where every processor exe- cutes the same program, but on different data. Here, load balancing is often reduced to the problem of balancing the data properly [3]. With the i