Supporting Malleable Parallel Applications on Clusters of Workstations

工作站集群越来越多地被用作并行计算的平台。在工作站集群上执行的并行应用程序必须具有容错能力，才能经受住网络和工作站故障，并且必须具有在检测到所有者活动时退出工作站的能力。可延展的并行应用程序还具有动态缩小或扩展其并行度的能力。初步研究表明，由于底层资源的暂时性，该属性在集群环境中会带来显著的性能优势。所提出的研究涉及设计和实现一个在工作站集群上支持可伸缩并行应用的执行环境。将为三类重要的并行应用程序设计和实现特定于领域的重构机制，这些机制可以改变程序的并行度，特别是松散同步的结构化网格应用程序、松散同步的非结构化网格应用程序和运筹学领域的一类异步不规则应用程序。这些机制将集成到支持这些应用程序的专用运行时库中。最后，将开发用于最小化动态重新配置开销的算法和技术。

Clusters of workstations are increasingly being used as a platform for parallel computing. Parallel applications executing on workstation clusters must be fault-tolerant in order to survive network and workstation failures, and must have the ability to withdraw from a workstation on detecting owner activity. Malleable parallel applications also have the ability to dynamically shrink or expand their degree of parallelism. Preliminary research indicates that this property leads to significant performance advantages in cluster environments because of the transient nature of the underlying resources. The proposed research involves the design and implementation of an execution environment that supports malleable parallel applications on clusters of workstations. Domain-specific reconfiguration mechanisms that enable changes in the degree of parallelism of a program will be designed and implemented for three important classes of parallel applications, specifically loosely synchronous structured grid-oriented applications, loosely synchronous unstructured grid-oriented applications, and a class of asynchronous irregular applications from the domain of operations research. These mechanisms will be integrated into specialized run-time libraries that support these applications. Finally, algorithms and techniques for minimizing the overhead of dynamic reconfiguration will be developed.