Architecture Specific Communication Optimizations for Structured Adaptive Mesh-refinement Applications

OF THE THESIS Architecture Specific Communication Optimizations for Structured Adaptive Mesh-Refinement Applications by Taher Saif Thesis Director: Professor Manish Parashar Dynamic Structured Adaptive Mesh Refinement (SAMR) techniques for solving partial differential equations provide a means for concentrating computational effort to appropriate regions in the computational domain. Parallel implementations of these techniques typically partition the adaptive heterogeneous grid hierarchy across available processors, and each processor operates on its local portions of this domain in parallel. However, configuring and managing the execution of these applications presents significant challenges in resource allocation, data-distribution and load balancing, communication and coordination, and runtime management. Due to their irregular load distributions and communication requirements across levels of the grid hierarchy, parallel SAMR applications make extensive use of non-blocking MPI primitives to reduce synchronization overheads. The behavior and performance of MPI non-blocking message passing operations are particularly sensitive to implementation specifics as they are heavily dependant on available system resources. As a result, naive use of these operations without an understanding of architectural constraints and the underlying MPI implementation can result in serious performance degradations, often producing synchronous