Improving
the Performance of Large-Scale Linear System Solution in
Distributed Computing Environments
Dr. Masha Sosonkina,
Abstract:
The
size and complexity of linear systems arising in modern high- performance
scientific applications present a computational challenge and require efficient
parallel solution techniques. In this talk, I consider iterative methods for
solving large-scale linear systems in distributed computing environments. For
these platforms, iterative solution methods is an
attractive choice because they exhibit a high degree of parallelism and have
moderate storage requirements. However, the questions of fast iterative
convergence, parallel efficiency, and workload balancing are still to be
answered.
First, I will describe a few linear system transformations that may improve
convergence while achieving good scalability characteristics. Then I will
present dynamic adjustments of solution parameters based on runtime communication
channel and compute node conditions. A more balanced parallel execution, which
leads to a better iterative convergence, is observed due to chosen adaptations.