Fine-Grained Task Scheduling Using Adaptive Data Structures

Task pools have been shown to provide efficient load balancing for irregular applications on heterogeneous platforms. Often, distributed data structures are used to store the tasks and the actual load balancing is achieved by task stealing where an idle processor accesses tasks from another processor. In this paper we extent the concept of task pools to adaptive task pools which are able to adapt the number of tasks moved between the processor to the specific execution scenario, thus reducing the overhead for task stealing significantly. We present runtime experiments for different applications on two execution platforms.