PDS: Architecture, Algorithms and Applications for Future Generation Supercomputers

本研究提出了100兆次浮点运算级计算机使能技术的初步评估。该研究大致分为以下几个部分：架构（单节点架构、并行架构）、系统软件（语言、编译器、编程工具和环境）、算法（设计模型、性能和可扩展性）和应用（可扩展性和可实现性能）。下一代计算机的单节点架构将利用相当大的隐式并行性。单节点架构的候选技术包括超标量执行、很长的指令字、并发硬件多线程和内存处理器。pi提出了一种基于VLIW的处理器体系结构，并可能使用pi实现。并行架构是基于这些节点在分层网络中的集群。由于计算单元数量庞大，算法设计和可扩展性问题变得至关重要。有必要对问题进行分层分解，并强调与底层体系结构的适当映射。较高层次上的问题必须是松散耦合的，而较低层次上的问题可以是紧密耦合的。必须构建替代算法设计方法和抽象机器模型。最后，需要确定能够有效使用这些体系结构的应用程序的范围。这项研究将对未来一代计算机的这些方面进行研究。pi希望从他们设计可扩展并行算法和架构以及CEDAR项目的经验中汲取大量经验。

This research proposes a preliminary evaluation of enabling technologies for 100 TeraFLOP class computers. The research is broadly classified into the following parts: architectures (single node architecture, parallel architecture), system software (languages, compilers, programming tools and environments), algorithms (design models, performance and scalability) , and applications (scaling and realizable performance). Single node architectures for future generation computers will exploit considerable implicit parallelism. Candidate technologies for single node architectures include super-scalar execution, very long instruction words, simultaneous hardware multithreading, and processors-in-memory. The PIs propose a processor architecture based on the VLIW with possible implementations using PIMs. Parallel architectures are based on clusters of these nodes in a hierarchical network. Due to the large number of computational units, algorithm design and scalability issues become critical. Hierarchical decomposition of problems with an emphasis on proper mapping to the underlying architecture is necessary. Whereas problems at higher levels must be loosely coupled, they can be very tightly coupled at lower levels. Alternate algorithm design methodologies and abstract machine models must be constructed. Finally, the scope of applications that can effectively use these architectures needs to be identified. This research will perform a study of each of these aspects of future generation computers. The PIs expect to draw heavily from their experience in the design of scalable parallel algorithms and architectures and from the CEDAR project.