Insulating the scientific programmer from perilous parallel architecture

The resource requirements of modern-day scientific applications is making hardware acceleration a requisite design consideration. But typically these applications are created and implemented by scientists trained in other fields, not the intricacies of parallel processing. Consequently, there is a definite need for abstractions. In this paper, we discuss why the typically object-oriented abstractions that are commonly made available in such settings introduce accidental sequentiality and render the hardware acceleration ineffective. We hence propose a few research directions that we intend to pursue in order to generate new abstractions that better insulate the programmer from implementation details that do not relate to application logic.