An Error Resilience Strategy of a Complex Moment-Based Eigensolver
An Error Resilience Strategy of a Complex Moment-Based Eigensolver
复制标题
DOI：
10.1007/978-3-319-62426-6_1
复制
发表时间：
2015-09
期刊：
影响因子：
--
通讯作者：
A. Imakura;Y. Futamura;T. Sakurai
中科院分区：
文献类型：
--
作者：
A. Imakura;Y. Futamura;T. Sakurai
最近，在高度并行环境中积极开发了基于复杂矩的特征求解器，以解决大型且稀疏的特征值问题。在本文中，我们提供了一种瑞利-里茨型复杂基于矩的并行特征求解器的错误恢复策略，用于解决广义特征值问题。我们的策略基于在发生位翻转等软错误的情况下特征求解器的错误界限。使用错误界限，我们实现了特征求解器固有的错误恢复能力，在最耗时的部分不需要标准检查点和复制技术。
Recently, complex moment-based eigensolvers have been actively developed in highly parallel environments to solve large and sparse eigenvalue problems. In this paper, we provide an error resilience strategy of a Rayleigh–Ritz type complex moment-based parallel eigensolver for solving generalized eigenvalue problems. Our strategy is based on an error bound of the eigensolver in the case that soft-errors like bit-flip occur. Using the error bound, we achieve an inherent error resilience of the eigensolver that does not require standard checkpointing and replication techniques in the most time-consuming part.