high performance computing on graphics processing units: hgpu.org

Scientific applications strive for increased memory and computing performance, requiring massive amounts of data and time to produce results. Applications utilize large-scale, parallel computing platforms with advanced architectures to accommodate their needs. However, developing performance-portable applications for modern, heterogeneous platforms requires lots of effort and expertise in both the application and systems domains. This is more relevant for unstructured applications whose workflow is not statically predictable due to their heavily data-dependent nature. One possible solution for this problem is the introduction of an intelligent Domain-Specific Language (iDSL) that transparently helps to maintain correctness, hides the idiosyncrasies of low-level hardware, and scales applications. An iDSL includes domain-specific language constructs, a compilation toolchain, and a runtime providing task scheduling, data placement, and workload balancing across and within heterogeneous nodes. In this work, we focus on the runtime framework. We introduce a novel design and extension of a runtime framework, the Parallel Runtime Environment for Multicore Applications. In response to the ever-increasing intra/inter-node concurrency, the runtime system supports efficient task scheduling and workload balancing at both levels while allowing the development of custom policies. Moreover, the new framework provides abstractions supporting the utilization of heterogeneous distributed nodes consisting of CPUs and GPUs and is extensible to other devices. We demonstrate that by utilizing this work, an application (or the iDSL) can scale its performance on heterogeneous exascale-era supercomputers with minimal effort. A future goal for this framework (out of the scope of this thesis) is to be integrated with machine learning to improve its decision-making and performance further. As a bridge to this goal, since the framework is under development, we experiment with data from Nuclear Physics Particle Accelerators and demonstrate the significant improvements achieved by utilizing machine learning in the hit-based track reconstruction process.