SHF: Small: PAW: Novel Functionality in Programming Models to Productively Abstract Wavefront Parallel Pattern

点击翻译按钮获取中文摘要

With the rapid and globally competitive development of faster computing systems that can compute up to one quintillion floating-point operations/second, it becomes imperative to update the computer programs that direct how data is analyzed. However, it has been a challenge for established and time-tested legacy scientific code, filling up hundreds to thousands of lines of code, to adapt and alter to exploit the rich computing capacity of these systems. This is largely a manpower issue as the adaptation of codes requires application developers to constantly re-write their program codes. The steep learning curve associated with both the intricacies of hardware and the ever evolving programming languages puts pressure on the developers and impedes the progress of science. The biggest challenge developers face is the inability to maintain a "write-once reuse multiple times" software. With all eyes on the development of an exascale machine - one that can compute data at the speed of the human brain - it is imperative to address this fundamental challenge. The aim of this project is to design high-level abstractions that can adapt scientific code to current and upcoming systems in a manner that enhances the performance of these machines, thus ensuring that these "fast-as-the human-brain" systems are flexible and adaptable enough to encourage the broader scientific community. The goal of this project to enable high performance, memory-efficient, portable and productive software framework for parallelizing complex parallel patterns such as 'wavefronts', commonly found in large scientific applications such as neutron radiation transport, bioinformatics and atmospheric science. To achieve this goal, the investigator is addressing critical performance portable questions at the algorithmic-level, programming framework-level and at the software design level. The project studies the applicability of well-explored polyhedral transformation frameworks along with task-based environments on novel hardware systems, importantly on pre- and upcoming exascale systems. The studies are also suggestive of shortcomings in current programming models paving the way to developing novel insights towards high-level software abstractions for multi-use in different/diverse projects simultaneously.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.