Optimizing Resource Allocation for DAGs of Gang Tasks on Heterogeneous Platforms

Real-time applications are increasingly compute- and data-intensive, often incorporating machine learning and AI workloads. These applications typically execute on heterogeneous platforms that combine multicore processors with GPUs and, in some cases, additional hardware accelerators. While gang task models effectively capture the computation on GPUs and other accelerators, Directed Acyclic Graphs (DAGs) are well-suited to represent the dependencies among different application components running across the multicore platform and accelerators. In this paper, we study DAGs of gang tasks scheduled under a federated approach on heterogeneous platforms. We introduce a method to optimize the allocation of cores and streaming multiprocessors (SMs) for each DAG. Our results show that the proposed method significantly reduces resource requirements per application compared to existing state-of-the-art approaches.