Framework for Extensible, Asynchronous Task Scheduling (FEATS) in Fortran

Most parallel scientific programs contain compiler directives (pragmas) such as those from OpenMP [1], explicit calls to runtime library procedures implementing the Message Passing Interface (MPI) [2], or compiler-specific language extensions provided by CUDA [3]. By contrast, recent Fortran standards empower developers express algorithms without directly referencing lower-level programming models [4, 5]. Fortran’s features place within Partitioned Global Address Space (PGAS) class of models. When writing that exploit data-parallelism, application often find it straightforward develop custom algorithms. Problems involving complex, heterogeneous, staged calculations, however, pose much greater challenges. Such applications require careful coordination tasks in a manner respects dependencies prescribed directed acyclic graph. rolling one’s own solution proves difficult, extending customizable framework becomes attractive. The paper presents design, implementation, and use Framework for Extensible Asynchronous Task Scheduling (FEATS), which we believe be first task-scheduling tool written modern Fortran. We describe benefits compromises associated with choosing implementation language, propose ways future can best support case this paper.