一类非张量积区域快速傅立叶变换算法在匡产并行机上的可扩展性测试

本文给出了一类基于六边形非张量积区域上的广义离散快速傅立叶变换算法（HFFT）以及它在国产百万亿次超级计算机（曙光5000A）上的测试运行情况．文章介绍了该算法在曙光5000A上的大规模集群测试加速比和可扩展性特性，并通过分析，说明HFFT在国产超级计算机的大规模并行环境下拥有良好的可扩展性．在使用8192个处理器核的情况下，HFFT加速比达到了277倍．我们同样对FFTw软件包进行了测试．本文的分析为解决其他科学计算程序在国产百万亿次规模集群上的可扩展性问题，提供了一些先行的参考和建议．

SCALABILITY EVALUATION OF HI-DIMENTIONAL FAST FOURIER TRANSFORMS ON DOMESTIC SUPERCOMPUTER

In this article we propose a set of fast algorithms for computing the HFFT （Hi-dimensional Fast discrete generalized Fourier Transforms） over the hexagon domains and the parallel dodecahedron domains. We evaluate the performance characteristic of HFFT on domestic 100Tera-Scale cluster （The Dawning 5000A Supercomputer）. The evaluation result indicates that HFFT algorithm scales well on Dawning 5000A. As using 8192 cores in the cluster, the speedup of HFFT is about 277. By using a specific recursive data structure, the scalability of the HFFT program is promising in this evaluation. We also evaluated the FFTW on the Dawning 5000A. The analysis in this article provides some useful advice and suggestions on the scalability of other scientific applications on the domestic 100Tera-Scale Supercomputer.