high performance computing on graphics processing units: hgpu.org

This research addresses node-level scalability, portability, and heterogeneous computing challenges facing asynchronous many-task (AMT) runtime systems. These challenges have arisen due to increasing socket/core/thread counts and diversity among supported architectures on current and emerging high-performance computing (HPC) systems. This places greater emphasis on thread scalability and simultaneous use of diverse architectures to maximize node use and is complicated by architecture-specific programming models. To reduce the exposure of application developers to such challenges, AMT programming models have emerged to offer a runtime-based solution. These models overdecompose a problem into many fine-grained tasks to be scheduled and executed by an underlying runtime to improve node-level concurrency. However, task execution granularity challenges remain, and it is unclear where and how shared memory programming models should be used within an AMT model to improve node use. This research aims to ease these design decisions with consideration for performance portability layers (PPLs), which provide a single interface to multiple shared memory programming models. The contribution of this research is the design of a task scheduling approach for portably improving node use when extending AMT runtime systems to many-core and heterogeneous HPC systems with shared memory programming models. The success of this approach is shown through the portable adoption of a performance portability layer, Kokkos, within Uintah, a representative AMT runtime system. The resulting task scheduler enables the scheduling and execution of portable, fine-grained tasks across processors and accelerators simultaneously with flexible control over task execution granularity. A collection of experiments on current many-core and heterogeneous HPC systems are used to validate this approach and inform design recommendations. Among resulting recommendations are approaches for easing the adoption of a heterogeneous MPI+PPL task scheduling approach in an asynchronous many-task runtime system and furthermore to ease indirect adoption of a performance portability layer in large legacy codebases.