Recently, manycore and multicore based massively parallel processing architectures, led by GPUs and multicore CPUs, have become the powerful and cost-effective hardware trend in the high performance computing landscape. To incorporate this processing power, various highly efficient parallel algorithms to exploit multilevel parallelism are being designed and developed. Among them, a novel combined space-time parallel approach that uses the parareal algorithm for time parallelization, demonstrates the possible multiplied speedup on top of that achieved with spatial parallelization. This approach is implemented in the EXN/Aero parallel Computational Fluid Dynamics (CFD) solution, developed by Envenio inc. and UNB CFD lab. Notably, the parareal performance with respect to the speedup and convergence when applied into the combined approach, is further complicated with the effects from the parallel spatial solver which hinders effective analysis. As a complementary tool to traditional experimental performance analysis with respect to speedup and convergence, in this thesis we propose a visual analysis platform, called EXN/Vizer, to help better understand and gain valuable insights into the parareal performance in search for potential optimization strategies. EXN/Vizer is designed as a modular and comprehensive GUI application platform based on the NetBeans Platform technology. Within EXN/Vizer, two layered major visualization modules are implemented using JavaFX toolkit along with several external utility library modules. The bottom layer visualization module is the EXNMonitor Trackboard for dynamically monitoring CFD solver control data including the CFD convergence data and field monitoring point data, which is provided by the CFD application as result datasets. The top layer module is the EXNParareal Dashboard as the general visualization framework for parareal solution process and performance analysis with the parareal profiling datasets. The live tests with EXN/Aero CFD simulations have shown that EXN/Vizer is highly efficient while performing interactive visualization tasks upon heavy dataset load of several thousands of data points, with the runtime heap memory consumption around 100MB at most and 60MB on average. In addition, the simulation solver control trackboard module is constructed for visualizing any SQLite database structured datasets and is also generalizable for other types of database structured datasets.
Zircon - This is a contributing Drupal Theme Design by WeebPal.
