Efficient Compilation Issues for Scalable Distributed-Memory Multicomputers

分布式内存，并行多计算机可以提供解决大挑战计算科学问题所需的高水平的性能。 在成本和可伸缩性方面，多计算机比共享内存多处理器具有显著的优势. 不幸的是，从这些机器中提取所有的计算能力需要用户为它们编写高效的软件，这是一个极其费力且容易出错的过程。 在分布式存储系统中，数据在处理器之间的分布对于并行程序的效率至关重要。在这个项目中，正在研究在这种机器中自动数据分配的问题。该方法是独特的，因为它是基于复杂的通信和计算的成本模型，由不同的目标机器经验测量的架构指标参数化。使用成本估计，数据分布被选择为通过最大化并行性（同时保持负载平衡）和最小化通信开销量（通过最大化数据局部性）来最小化程序的总体执行时间。静态和动态分布都将在一个统一的框架中得到支持，以自动选择可以在程序执行过程中动态变化的数据分布，从而为大型复杂应用程序提供可扩展的并行性能。这种方法不仅具有坚实的理论基础，而且正在集成到一个复杂的编译器中，该编译器实际上可以为支持所有块，循环和块循环数据分布的可变数量的处理器生成代码。在这个项目中调查的编译技术也将通过一个新的基于间隔的表示与并行应用程序中的定期和不定期访问的同时支持集成。 ***

Distributed-memory, massively-parallel multicomputers can provide the high levels of performance required to solve the Grand Challenge computational science problems. Multicomputers offer significant advantages over shared- memory multiprocessors in terms of cost and scalability. Unfortunately, extracting all the computational power from these machines requires users to write efficient software for them, which is an extremely laborious and error-prone process. The distribution of data across processors is of critical importance to the efficiency of the parallel program in a distributed memory system. In this project the problem of automated data distribution in such machines is being investigated. The approach is unique in that it is based on sophisticated cost models of communication and computation that are parameterized by architectural metrics empirically measured for different target machines. Using the cost estimates, data distributions are selected to minimize the overall execution time of the program by maximizing parallelism (while maintaining load balance) and minimizing the amount of communication overhead (by maximizing data locality). Both static and dynamic distribution will be supported in a unified framework to automatically select data distributions which can dynamically change over the course of a program's execution in order to provide scalable parallel performance for large, complex applications. This approach not only has a solid theoretical basis, but is being integrated into a sophisticated compiler which can actually generate code for a variable number of processors supporting all block, cyclic, and block-cyclic data distributions. The compilation techniques investigated in this project will also be integrated with simultaneous support for regular and irregular accesses in parallel applications through a novel interval-based representation. ***