Patterns for OpenMP Task Data Dependency Overhead Measurements

Starting with version 4.0, the OpenMP standard has introduced data dependencies to provide a way for synchronizing the concurrent execution of task based on dataflow information. This indirect approach to fine-grained sychronization offers a convenient way for creating a task graph without having to explicitly synchronize individual tasks and can be used to parallelize both regular and irregular applications to expose a higher level of concurrency to the runtime system. However, the cost associated with task creation and management, including matching input and output dependencies, is a crucial factor in designing the granularity of individual tasks, i.e., the amount of work to encapsulate in a task. In this work, we present a set of benchmarks designed to determine the overhead associated with dependency management and give an overview of the performance characteristics of a set of compilers widely used in parallel computing. We hope to provide application developers with a way to make informed decisions on the granularity of their tasks given the dependency patterns dictated by the algorithm. Our benchmark results show varying performance characteristics of different implementations that are both interesting and important to have in mind throughout the task design process.