Techniques for Finite Element Methods on Modern Processors

In this thesis, methods for efficient utilization of modern computer hardware for numerical simulation are considered. In particular, we study techniques for speeding up the execution of finite-element methods. One of the greatest challenges in finite-element computation is how to efficiently perform the the system matrix assembly efficiently in parallel, due to its complicated memory access pattern. The main difficulty lies in the fact that many entries of the matrix are being updated concurrently by several parallel threads. We consider transactional memory, an exotic hardware feature for concurrent update of shared variables, and conduct benchmarks on a prototype processor supporting it. Our experiments show that transactions can both simplify programming and provide good performance for concurrent updates of floating point data. Furthermore, we study a matrix-free approach to finite-element computation which avoids the matrix assembly. Motivated by its computational properties, we implement the matrix-free method for execution on graphics processors, using either atomic updates or a mesh coloring approach to handle the concurrent updates. A performance study shows that on the GPU, the matrix-free method is faster than a matrix-based implementation for many element types, and allows for solution of considerably larger problems. This suggests that the matrix-free method can speed up execution of large realistic simulations.