| Application codes in computational science and engineering are concerned largely with simulating physical phenomena, which have complicated data, geometry and communication requirements. Dynamically adaptive techniques seek to enhance the initial discretization of the computational domain and the numerical methods by applying a finer resolution only in the regions that require higher resolution. In such simulations, the per-processor work load or the geometric locality of objects changes as the computation proceeds so that an initially balanced decomposition is then either unbalanced or has unacceptably high amounts of communication.; Due to the complexity as well as the large problem size of adaptive parallel applications, the development of appropriate tuning schemes with regard to domain partitioning and repartitioning is still challenging. Traditional profiling techniques for tracing execution behaviors are not suitable for this purpose, because they cannot provide a spatial view of application behavior associated to the underlying computation domain/mesh.; The work presented in this thesis is based on the belief that fine-grained profiling of mesh-based applications mapping to the mesh structure could provide a more complete view of the application performance. Thus, unlike traditional profiling approaches which target the tracing of events or determining performance parameters for subroutines, the approaches described here attempt to discover the inherent adaptivity of parallel applications mapped to the computation domain to aid in the performance tuning of parallel applications, especially with dynamic load balancing. Based on the profiling results, we also designed a simulation framework which is driven by a sequence of application traces mapping to the mesh structure. By integrating Load balancing software into the system, this framework can provide a convenient test bed for understanding application adaptivity and developing optimization schemes.; Our profiling schemes only require one-time execution of the target program on any platform to generate a sequence of trace with timestamps. The traces can then be fed to the simulator for experiments under various system configurations, independent of the real application. Experiments have been performed to verify the proposed profiling techniques and trace-driven simulations. |
