Flexible framework for commodity FPGA cluster computing

With falling costs and successful demonstrations of performance, FPGAs have become a prime candidate for use in high performance computing. However, the use of FPGA technology in clustered environments has largely been limited to commercial and/or proprietary designs that require developers to learn new programming models and software tools. In this paper, a framework is presented which enables development of parallel programs across application-specific reconfigurable hardware using simple hardware interface abstractions and standard MPI application structure. This approach leverages commodity technologies including embedded Linux running on hardwired PowerPC processors to manage communication such that each hardware acceleration unit can function as a fully MPI-2 compatible node. The implied hardware/software design and programming model are discussed and an application case study is presented to demonstrate functionality and elucidate platform benefits through performance analysis.