Multithreading for Dynamic Unstructured Grid Applications

摘要

The success of parallel computing in solving realistic computational applications relies on their efficient mapping and execution on large- scale multiprocessor architectures. When the algorithms and data structures corresponding to these problems are unstructured or dynamic in nature, efficient implementation on parallel machines offers considerable challenges. Unstructured applications are characterized by irregular data-access patterns whereas dynamic mesh adaptation causes computational work- loads to grow or shrink at run time. For such applications, dynamic load balancing is required in order to achieve algorithmic scaling on parallel machines. Our objectives were to implement various parallel versions of a dynamic unstructured algorithm and to critically compare their performances in terms of run time, scalability, programmability, portability, and memory overhead.