Declarative Data Flow in a Graph-Based Distributed Memory Runtime System

Abstract Runtime systems can significantly reduce the cognitive complexity of scientific applications, narrowing gap between engineering and domain science in HPC. One most important angles this is automating data migration a cluster. Traditional approaches require application developer to model communication explicitly, for example through MPI primitives. Celerity, runtime system accelerator clusters heavily inspired by SYCL programming model, instead provides purely declarative approach focused around access patterns. In addition eliminating need explicit transfer operations, it basis efficient dynamic scheduling at runtime. However, currently only suitable accessing array-like from runtime-controlled tasks, while real programs often interact with opaque local each host, such as handles or database connections, also defined way transporting into out virtualised buffers paper, we introduce graph-based API expressing side-effect dependencies tasks moving context space.