KFusion: optimizing data flow without compromising modularity

Abstract

Programming language support for multi-core architectures introduces a fundamentally new mechanism for modularity---a kernel. Though it can be used as a means to separate concerns, a kernel is given a clean slate of memory at execution time. As a consequence, application developers attempting to leverage libraries of kernels often incur substantial unanticipated performance penalties. Currently, the only recourse is to compromise modularity for the sake of optimizing data flow on an application-specific basis. KFusion is our prototype tool for optimizing libraries of kernels according to application-specific needs. Our goal is to shield application developers from loop fusion and deforestation in compositions of low level kernels that share data. Libraries, augmented by domain experts with annotations to ensure correct compositions of kernels, provide application developers with the opportunity to supply hints according to customized data flow needs---keeping modularity intact. In the worst case, an inaccurate hint incurs no penalty. Case studies of applications using general-purpose libraries for linear algebra, image manipulation and physics engines show that KFusion can substantially improve performance associated memory bandwidth bottlenecks.