Reservoir Labs, Inc. — Department of Energy SBIR Phase II: 03b

Description

Achieving exascale is critical for improving America’s economic competitiveness and making scientific breakthroughs. The road to exascale is evolving toward advanced computer architectures with diverse processors. Advanced software solutions on programming models and runtimes are needed to obtain the best performance from these new hardware. The advanced programming paradigms are however complex, impose verbose expressions of algorithms, and demand expensive and rare expertise. The overall objective of the project is to develop an automatic optimization tool for mapping programs to exascale systems using advanced task-based runtime models that (1) provides high productivity and portability, (2) delivers high-performance compilation for heterogeneous platforms by supporting advanced task-based runtimes targeting heterogenous execution and/or coordinating multiple task-based runtimes to co-exist and operate on heterogeneous platforms, and (3) enables performance portability while automatically mapping programs to exascale systems. In Phase I of the X-MAT project, we extended an existing parallelizing compiler with a new generic task-based runtime layer and instantiated the generic runtime layer with several task-based runtimes including OpenMP, Legion, and a custom task runtime for GPGPUs. We implemented cross-runtime interactions required to run multiple task runtimes simultaneously, allowing support for hierarchical parallelism and/or heterogeneity. In Phase II of the X-MAT project, we will broaden and deepen the compiler support for automatic mapping to task-based runtime backends for guaranteed exascale benefits, by adding more runtimes (namely, Kokkos, DARMA, PaRSEC, CUDA graphs) and hardening the runtimes introduced in Phase I. We will add support to provide visibility into compiler-generated transformations for task-based parallelism for enhancing performance debugging and porting. We will improve and demonstrate the application scope of task-based runtime mapping with increased robustness. Commercial Applications and Other Benefits: The performance and productivity benefits offered by X-MAT will enable DOE and the users of exascale ecosystem to push the frontiers of science and technology to enable scientific discovery and enhance national security. This is critical for maintaining the global leadership of US economy. During the Phase II timeframe and beyond, the Government labs will have large installations of new machines (e.g., CORAL2) and will be striving to maximize the utilization of the Leadership Computing Facilities. A mature product built using the X-MAT technology will be useful to the users mapping their applications to such machines. Commercial sectors (e.g. financial sector, biosciences industry, and the oil and gas industry) face the need to port their applications or write new applications to modern powerful computing systems. X-MAT will offer tangible benefits to these commercial sectors.