Benchmark cases for linear systems and scalability analysis
Posted: 04 May 2023, 13:51
Hi All,
We will be doing some comparisons of different linear solver strategies for CPU time and scalability analysis over the summer. If you have some ideas what problems deserve to be analyzed we are open for input.
What problem should we study?
We are also open for suggestions for test cases. The should be scalable and simple in the sense that only one equation is analyzed at a time. The focus is more on the linear system level so the cases could ideally be rather canonical. Scalable test require that the mesh can be easily obtained in different densities. This may mean internal mesh multiplication by using "Mesh Levels" or easily refined meshes with some open source tools. Probably the biggest tests will be require mesh multiplication for convenience.
The size of the problem may vary. We are limited by 32-bit integers to 2,147,483,647 dofs (2e9) and by queuing system of the supercomputer to 51,200 processes. The easiest and most scalable problems (Poisson with MG) could reach that but more challenging problems won't certainly be this large.
As the desired outcome of this work is some processes that could help people to choose the most efficient linear solver strategy. Also the scalability results of the test problems will be published.
-Peter
We will be doing some comparisons of different linear solver strategies for CPU time and scalability analysis over the summer. If you have some ideas what problems deserve to be analyzed we are open for input.
What problem should we study?
We are also open for suggestions for test cases. The should be scalable and simple in the sense that only one equation is analyzed at a time. The focus is more on the linear system level so the cases could ideally be rather canonical. Scalable test require that the mesh can be easily obtained in different densities. This may mean internal mesh multiplication by using "Mesh Levels" or easily refined meshes with some open source tools. Probably the biggest tests will be require mesh multiplication for convenience.
The size of the problem may vary. We are limited by 32-bit integers to 2,147,483,647 dofs (2e9) and by queuing system of the supercomputer to 51,200 processes. The easiest and most scalable problems (Poisson with MG) could reach that but more challenging problems won't certainly be this large.
As the desired outcome of this work is some processes that could help people to choose the most efficient linear solver strategy. Also the scalability results of the test problems will be published.
-Peter