Significantly improving the scalability of large structured adaptive mesh refinement (SAMR) applications is challenging. It requires sophisticated capabilities for using the underlying parallel computer's resources in the most efficient way. This is non-trivial, since the basic conditions for how to allocate the resources change dramatically during run-time due to the dynamics inherent in these applications.
This report presents a first characterization of a hybrid and dynamic partitioner for parallel SAMR applications. Specifically, we determine optimal parameter settings for trade-offs like communication vs. load balance and speed vs. quality. The key contribution is that the characterization enables the partitioner to respond accurately to stimuli from system and application state, and hence adapting to various SAMR scenarios. This potentially reduces run-time for large SAMR applications.
Available as Postscript (815 kB), compressed Postscript (268 kB), and PDF (338 kB)
Download BibTeX entry.