Distributed-Memory Multiprocessors in FPGAs

The exploitation of parallelism in general purpose soft-core processors has been increasingly considered an efficient approach to accelerate embedded applications. Therefore, it’s important to use standard parallel programming paradigms that facilitate the development of parallel applications, abstracting the user from architectural details. The Message Passing Interface (MPI) is a standard library to develop messagepassing programs for distributed memory processing systems. This work proposes a Message Passing Interface for FPGA soft-processors and Zynq heterogeneous systems. The work included the definition of a fully functional set of MPI functions, which has been developed as a portable C library, and the design of a set of configurable hardware components to support the communication between all the processors. Considering the specifics of the target devices, namely the resource limitations in comparison with supercomputers or clusters of workstations, the design emphasized low resource utilization as well as hardware scalability and software reliability. A set of benchmarks covering a wide range of algorithms was used to evaluate the work developed. The experimental results fully validated the implemented designs and showed that standard MPI applications can be easily ported to the target platforms. Maximum efficiencies (up to 100%) were achieved for the algorithms with lower communication overheads, such as the cpi for pi calculus. Keywords—Parallel Computing, High-Performance Computing, Embedded Systems, Soft-Processors, FPGAs, MicroBlaze, Zynq, MPI