Minimal MPI as programming interface for multicore System-on-Chips

The decrease of feature sizes allows integrating a growing number of functional blocks into a System-on-Chip (SoC). This allows to include multiple general purpose CPU cores into SoCs, so that applications can be parallelized to achieve higher performance. To provide communication between the cores of such multicore SoCs, different application programming interfaces (APIs) have emerged. In this paper, we propose to use a message-passing interface as basic API, since it makes low demands to the hardware while providing high flexibility and scalability, along with a lightweight, minimal implementation. We show the feasibility of this approach by a parallel implementation of a scientific application that is based on the proposed API. Evaluation on Intel's Single-Chip Cloud Computer, a research chip that features 48 processor cores on a single chip, shows that such a lightweight implementation provides higher speedup and better scaling properties than standard message-passing libraries.