Section: New Results

Hierarchicals Tasks

The programming model of StarPU, namely the sequential task flow model, was successfuly used in several applicative areas and was able to achieve high performance. However, the submission process needed either to be completely static, in the sense that the whole task graph is submitted at once, or to be stopped from time to time in order to control the execution. To overcome these limitations, we have introduced a new paradigm which we call hierarchical tasks where the so-called control tasks allow to submit at runtime a task subgraph. By allowing the submission of some parts of the task graph to be delayed until the execution of the corresponding control task, this feature allows to timely and dynamically choose the right version of the computation task subgraph (e.g. OpenMP, StarPU, cuda, or sequential, etc. implementations).

The graph of control tasks provide a high-level description of the computations which allow to use and design sophisticated scheduling algorithms. Furthermore, the cost of managing the control graph being much smaller than the one of the computation task graph, relying on the hierarchical tasks scheme enhances the scalability of the runtime system and allows to parallelize the submission process. Finally, this mechanism represents an elegant way of tackling the granularity issues which represent a key problem for achieving high performance in a heterogeneous context. The specificity of our implementation is to nicely combine hierarchical tasks with data partitioning, without needless synchronisation points.