Disputation: James D. Trotter
Doctoral candidate James D. Trotter at the Department of informatics, Faculty of Mathematics and Natural Sciences, is defending the thesis High-performance finite element computations: Performance modelling, optimisation, GPU acceleration & automated code generation for the degree of Philosophiae Doctor.
The University of Oslo is closed. The PhD defence and trial lecture will therefore be fully digital and streamed directly using Zoom. The host of the session will moderate the technicalities while the chair of the defence will moderate the disputation.
Ex auditorio questions: the chair of the defence will invite the audience to ask ex auditorio questions either written or oral. This can be requested by clicking 'Participants -> Raise hand'.
Main research findings
Computer experiments have become a valuable tool for investigating various physical and biological processes described by partial differential equations (PDEs), such as weather forecasting or modelling the mechanical behaviour of cardiac tissue. Finite element methods are a class of numerical methods for solving PDEs that are often preferred, but these methods are rather difficult to implement correctly, let alone efficiently.
This thesis investigates the performance of several key computational kernels involved in finite element methods. First, a performance model is developed to better understand sparse matrix-vector multiplication, which is central to solving linear systems of equations that arise during finite element calculations. Second, the process of assembling linear systems is considered through careful benchmarking and analysis of the memory traffic involved. This results in clear guidelines for finite element assembly on shared-memory multicore CPUs.
Finally, hardware accelerators are incorporated by extending the FEniCS PDE solver framework to carry out assembly and solution of linear systems on a graphics processing unit (GPU). Example problems show that GPU-accelerated finite element solvers can exhibit substantial speedup over optimised multicore CPU codes. Moreover, the use of automated code generation makes these techniques much more accessible to domain scientists and non-experts.
Contact information to Department: Anniken R. Birkelund