Parallel Divide-and-Conquer Algorithm for Solving Tridiagonal Eigenvalue Problems on Manycore Systems
Parallel Divide-and-Conquer Algorithm for Solving Tridiagonal Eigenvalue Problems on Manycore Systems
复制标题
求解众核系统上三对角特征值问题的并行分治算法
DOI：
10.1007/978-3-319-78024-5_54
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
Imamura Toshiyuki
中科院分区：
文献类型：
--
作者：
Hirota Yusuke;Imamura Toshiyuki
提出了一种基于批处理核执行调度的多核系统上求解实对称三对角特征值问题的并行分治（DC）算法。我们的算法具有更高的并行性，需要更少的全球同步比传统的算法。我们将基于我们算法的求解器的性能与英特尔MKL的DC求解器和PLASMA的求解器在Xeon E5、Xeon Phi Knights Corner和Xeon Phi Knights Landing上的性能进行了比较。数值测试表明，我们的算法的实现是可比的英特尔MKL至强E5和英特尔MKL和PLASMA的两个至强融核系统上。
We present a new parallel divide-and-conquer (DC) algorithm based on an execution scheduling by batched kernels for solving real-symmetric tridiagonal eigenvalue problems on manycore systems. Our algorithm has higher parallelism and requires less global synchronizations than a conventional algorithm. We compared the performance of the solver based on our algorithm with that of Intel MKL’s DC solver and PLASMA’s one on Xeon E5, Xeon Phi Knights Corner, and Xeon Phi Knights Landing. The numerical tests show that the implementation of our algorithm is comparable to Intel MKL on Xeon E5 and outperforms Intel MKL and PLASMA on the two Xeon Phi systems.