Data Parallel Programming on Diverse Architectures: Tools and Algorithms

本项目研究同步、数据并行方法在并行算法开发中的威力。使用这种方法，许多数值和非数值算法都可以用优雅的方式表达出来。本项目的重点是由Think Machines公司开发的数据并行语言C*。如果可以编写允许在连接机、多计算机和共享内存多处理器上高效执行并行算法的C*编译器，则将有一整类算法无需重新编程即可将应用程序从一个并行体系结构转移到另一个并行体系结构。初步结果令人鼓舞，PI正在为超级立方体多计算机实现优化的C*编译器。该项目的重点是通过提高编译程序的效率、实现交互式调试器、探索更多的数据并行算法和它们所需的优化以及将C*编程环境移植到不同的体系结构来进一步发展早期的成功。

This project investigates the power of the synchronous, data- parallel approach to parallel algorithms development. Many numerical and nonnumerical algorithms can be expressed in an elegant fashion using this methodology. This project focuses on C*, a data parallel language development by Thinking Machines Corporation. If C* compilers can be written that allow for the efficient execution of parallel algorithms on the Connection Machine, multicomputers, and shared memory multiprocessors, then there would be a whole class of algorithms for which no reprogramming would be necessary in order to move an application from one parallel architecture to another. Preliminary results are encouraging, and the PIs are in the process of implementing an optimizing C* compiler for hypercube multicomputers. The focus of this project is to build upon earlier successes by improving the efficiency of the compiled programs, implementing an interactive debugger, exploring more data parallel algorithms and the optimization they require, and porting the C* programming environment to different architectures.