SoC-MPI: A Flexible Message Passing Library for Multiprocessor Systems-on-Chips

Abstract

Current technologies allow designers to use several components, like processors, memories or hardware accelerators to build a multiprocessor system on-chip. In such systems, communication can be done by passing messages between components. In this paper, a flexible subset implementation of the message passing standard, called SoC-MPI, for multiprocessor systems on-chip is presented. The implementation hides the complexity of hardware from the programmer, but then uses hardware supported functions to enhance communication. Additionally, the message passing library is integrated in an architecture synthesis flow, which has the goal to create the best possible architecture from a given parallel application. Hence several communication networks (currently limited to Xilinx FPGAs) are supported by the library to enhance design flexibility. The SoC-MPI library is evaluated with several benchmarks, including the singular value decomposition.