STTR Phase I: Fault-Tolerant MPI: An Enabling Strategy, Product Concept, and Enhancement of Production Cluster Computing

这份小型企业技术转移研究（STTR）第一阶段的提案建议开发一个可扩展的、容错的和可靠的消息传递接口。MPI-2是未来几年在大规模并行机和Beowulf集群上实现可扩展并行计算的关键技术。在MPI-1和MPI-2标准中，容错是不存在的，并且没有令人满意的产品或研究结果提供了一个有效的路径，以提供生产可扩展计算应用程序具有有效的容错。该提案解决了MPI-1和MPI-2标准中的容错问题，并关注关键应用程序类、无故障开销和检查点重启策略。它将科学界和工业界中各种类型的MPI用户使用的资源管理基础设施与容错机制联系起来，以适用于实际的科学计算。如果成功，该产品将为并行程序、编程器和集群系统提供关键的新功能，包括增强基于MPI的现有商业应用程序，如CFD应用程序。所提出的努力是第一个实现容错MPI-2的技术，该技术将被科学家和工程师利用，因为基于MPI的并行计算是科学发现的广泛推动力。该项目预计将增加MPI-2的采用，以及集群和网格的更高生产力并行计算。通过这项工作获得的计算机科学和工程经验将使MPI实现和资源分配器能够更好地互操作，这反过来又将进一步实现高效的生产并行计算。针对流行的恢复机制，为关键类的应用程序的优化，可以应用于任何中间件，因此将导致提高应用程序的性能一般。

This Small Business Technology Transfer Research (STTR) Phase I proposal proposes to develop a scalable, fault-tolerant and reliable message-passing interface. MPI-2 is a key technology for the next several years in enabling scalable parallel computing on massively parallel machines and Beowulf clusters. Fault-tolerance is absent in both the MPI-1 and MPI-2 standards, and no satisfactory products or research results offer an effective path to providing production scalable computing applications with effective fault-tolerance. This proposal addresses fault-tolerance in both the MPI-1 and MPI-2 standards, with attention to key application classes, fault-free overhead, and checkpoint-restart strategies. It connects the resource management infrastructure in use by many types of MPI users in science and industry with the fault-tolerant mechanisms, to be useful in practical scientific computing. If successful this product will provide key new capabilities to parallel programs, programmers, and cluster systems, including the enhancement of existing commercial applications based on MPI, such as CFD applications. The proposed effort is the first towards realizing a fault-tolerant MPI-2, a technology that would be exploited by scientists and engineers across the spectrum, since parallel computing based on MPI is a widespread enabler of scientific discovery. The project expects to increase the adoption of MPI-2, as well as higher productivity parallel computing for clusters and potentially for the grid. The computer science and engineering experience gained through this work will enable better interoperation of MPI implementations and resource allocators, which in turn will further enable efficient production parallel computing. The optimizations targeted at the popular recovery mechanisms, for key classes of applications, can be applied to any middleware and hence would result in improving the performance of applications in general.