Prelegent: Dr Katarzyna Świrydowicz (Pacific Northwest National Laboratory, USA)
Data i godzina: 8 stycznia 2024 r., g. 10:30
Abstract: A need to solve an extremely large, sparse linear system commonly occurs in applications that are otherwise HPC-ready. While often overlooked, the repeated estimations of the solutions to linear systems can constitute ninety or more percent of the total application time. In the talk, I will share my experience gained while work on linear solvers developed during multiple subprojects of the Exascale Computing Project, which was a large scientific effort, spanning multiple US universities and national laboratories, and sponsored by the US Department of Energy. I will explain the limitations and challenges in the usage of linear solvers in applications. I will also present a variety of theoretical- and implementation-based techniques that have lead to reliable, GPU-ready linear solvers.