SHF:Small:Performance Portable Parallel Programming on Extremely Heterogeneous Systems

点击翻译按钮获取中文摘要

The computers that are deployed today are increasingly complex as their designers strive to increase the speed with which computations are performed, while simultaneously maintaining or even reducing their power consumption. Many of them include energy-efficient accelerator devices. Adapting existing application programs so that they can execute well on new computer systems where such devices are configured is a labor-intensive and error-prone activity that requires significant expertise. Moreover, unless portable standards are used, different versions of a program may need to be created for different hardware. The effort required to do so may delay, or even prevent, many codes from fully exploiting new systems. This project will learn how to effectively utilize Machine Learning methods to help automate the adaptation process. Specifically, it will learn how to modify applications that already run on multicore platforms so that they can effectively exploit accelerator devices. At the same time, it will study and develop best practices with respect to utilizing Machine Learning in the context of improving the performance of applications.This project will study and develop Machine Learning (ML)-based strategies and techniques to identify and extract code regions in technical applications that are suitable for mapping to the devices configured on a heterogeneous architecture. It will moreover develop the runtime technology needed to manage the execution of the resulting code. To accomplish this, the project will focus on application codes that have been parallelized to exploit multiple processing cores using the widely adopted, portable industry standard OpenMP and will use and extend features of the most recent OpenMP specification to express the device code and data mappings in a manner that is portable and permits subsequent manual optimization. The embedding of key choices in the code will aid performance portability. A key element of this research is the study of state-of-the-art ML methods, including classical ML and Deep-Learning techniques, with respect to their suitability for enhancing compilers and tools. An exploration of their relative merits for use in the compiler includes how to represent a compiler problem as a regression or classification problem. Research will also study approaches to code representation and the generation of sufficient data to train quality ML models. A set of benchmarks and mini-apps will be used to guide and evaluate the research. The project will participate in the work of the OpenMP Language Committee, will make practical results available via the open source LLVM infrastructure, will contribute to teaching and training materials, and will use this effort to enrich an ongoing collaboration with an HBCU.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.