Make software harder

ABSTRACTProgramming models should ideally let programmers express their ideas directly and explicitly. No such models are available for expert programmers, authors of machine-specific code optimizations, whose ideas often exploit architectural details invisible even at the assembly language level. In this paper we call for filling this void, and define performance transparency to evaluate the extent to which a programming model accommodates the expert programmer use case. We propose a specific approach to attaining performance transparency: make software harder by exposing the key components of the target architecture. The opportunities and challenges this approach brings are discussed.