Section: New Results

Static and Dynamic Adaptation of Task parallelism

This work is the result of Pierre Huchant PhD thesis, and the objectives are to adapt statically and dynamically OpenCL tasks. The adaptation consists in splitting tasks automatically into multiple sub-tasks, taking into account the heterogeneity of the architecture (sub-tasks are specifically created for one processing unit), the load imbalance within a parallel OpenCL, between the different iterations in space, and if the task graph is repeatedly executed, between the iterations in time, and it takes into account the time of the communications generated by splitting the tasks [2]. The method is able to cope with sequential task graphs (tasks are parallel themselves but scheduled sequentially) and deals with tasks manipulating complex data structures as shown on an N-body particle simulation mini-app.