Frau Prof. Gudula Rünger
"Task-based Parallel Programming in Scientific Computing"
14:00 Uhr, 1/336
Alle interessierten Personen sind herzlich eingeladen!
Abstract:Many applications from scientific computing exhibit an inherent modular structure of cooperating subtasks. For the execution of such applications on parallel or distributed platforms, a programming model using multiprocessor tasks is a suitable approach. Internally, a multiprocessor task can incorporate different kinds of parallelism or can consist of a hierarchical structure of smaller tasks. This multi-level parallelism can be exploited to achieve efficient programs.
In this talk, we discuss a programming environment and runtime system to support multiprocessor task programming based on a flexible specification of the task structure. Depending on the specific task structure of an application and the specific target platform, different functionalities are required. Statically known task structures may benefit from an a priori scheduling based on runtime predictions for computation and communiation. A dynamic scheduling is required for irregular applications or heterogeneous platforms.