DESC0024745

Heterogeneous computing architectures and large compute clusters can provide significant performance improvements on computationally intensive scientific workloads, but they require a great deal of domain expertise, time, and effort to use effectively that most scientific algorithm developers do not have. There are a number of software tools that address some of the individual aspects of porting and scaling an application to a heterogeneous compute cluster, but a single cohesive environment leveraging those tools is needed to significantly lower the barrier to entry of these tasks. A development environment will be created to enable scientists and engineers to construct their algorithms and easily map the parts of that algorithm to one or more of the heterogeneous computing elements available on their system. This will be achieved with a combination of existing distributed heterogeneous computing tools as well as transforms and tools developed as part of this effort. These transforms and tools will enable computation to be mapped to any node within a cluster, and present these interfaces to the user in a way that is easily understandable and requires no expert programming knowledge or experience. In Phase I, proof-of-concepts of core transforms will be developed, using nuclear physics experimental algorithms as exemplar use cases. A proof-of-concept distributed memory management system will deploy the exemplar algorithms across many nodes within a cluster automatically. Finally, an initial version of the development environment user interface will be designed to enable the user to structure, map, schedule, deploy, and run their application on a heterogeneous compute cluster. This project has the potential to dramatically lower the barrier to entry to supercomputing, enabling those without significant programming expertise to make use of heterogeneous computing resources, generate output data much quicker, and thus accelerate the pace of scientific computation. Beyond nuclear physics applications, the resulting computing architecture will allow scientists and researchers in academia, national labs and industry to more efficiently and effectively use high performance computing resources for their complex computational needs.