Distributed Task Parallel Mode Scheduler

The DTPM scheduler kernel enables applications focusing on a task parallel approach. In general, the scheduler follows a static scheduling approach, based on load balancing indicated by optional plugin weights. The execution of the individual plugins is distributed among the available MPI processes. Therefore, considerable speedup of the task execution can be achieved, if the task graph offers parallel paths. The following figure depicts the flow diagram of the scheduler. Note that the task graph is partitioned utilizing the METIS library, based on a load balancing approach guided by user-provided plugin weights. The tasks associated with the subgraphs are distributed to the available MPI processes, where they are executed.