Adaptive Auto-tuning Technology Aiming Complex Multicore and Multiprocessor Environments

为了在下一代超级计算机系统的多核处理器和多gpu上建立“自动调优技术”的关键组成部分，我们进行了这个研究项目。在这项工作中，我们开发了一种“利用GPU性能稳定的自动调优技术”和一种“利用性能数据库的自动调优技术”。此外，通过对环境条件动态变化时自动调谐的数学研究，成功地给出了基于贝叶斯模型的半最优实验设计。在集群系统或多核计算机系统上开发了特征值解算器、快速傅立叶变换和带预处理的迭代解算器的自动调谐。

To establish a key component of"automatic tuning technology"on multi-core processors and multi-GPU's for a next-generation supercomputer system, we conducted this research project. In this work, we developed an"automatic tuning technique employing performance stabilization on GPU's"and an"automatic tuning technique which utilizes a performance database". Furthermore, we successfully gave the semioptimal experimental design based on a Bayesian model through the mathematical research of the automatic tuning when the environmental conditions change dynamically. AT(automatic-tuning) for eigenvalue solvers, Fast Fourier Transform, and iterative solver with preconditioning were exploited on a cluster system or a multicore computer system.