Multi-core Portability Abstraction

Application portability between different multicore architecture-parallel programming paradigm/tool pairs is a big problem nowadays leading often to a complete rewrite of an application when switching from an architecture-paradigm pair to another. This is caused by a wide variety of architectural properties requiring different optimization techniques for different architectures, typically hiding the essence of (parallel) computation defined by the application. In this paper, we introduce the Multi-Core Portability Abstraction (MCPA) simplifying portability and implementation of parallel applications making use of shared memory. It abstracts away typical architecture dependent effects caused by latency, synchronization, and partitioning and acts as an executable intermediate abstraction/reference implementation as well as a tool for analyzing the intrinsic parallelism of the application and relative goodness of architectures in executing it. We give a short application example with performance measurements.