Section: New Results

Hierarchical Tasks

Modern computing platforms are heterogeneous and the load balancing is more complex to reach high performance. We decided to deal with the granularity problem in the context of task paralleliism and in a dynamic way through the implementation of hierarchical tasks in StarPU runtime. The idea is to give the runtime the ability to control tasks submission in order to choose the good granularity at the right moment. The application describes a control graph and the runtime generates the computation tasks graph on-the-fly according to the state of the machine (available computing resources, memory consumption, ...). As a consequence the runtime is able to limit the size of the computation tasks graph without loosing parallelism. Some experiments have been done on a Cholesky application and in the qr-mumps software and show that the work of an application programmer can be alleviated and the granularity choice could be easily delegated to the task based runtime.