Implementation of an automatic program partitioner on a homogeneous multiprocessor

The design and the first results of a prototype multiprocessor featuring the automatic partitioning of Fortran programs are presented. The conversion of the source program to a task schedule is based on a dataflow analysis that takes into account the task size, the number of processors, and the communication between tasks. Various levels of task granularity can be selected, which allows a tradeoff between the amount of parallelism and the communication overhead. The architecture consists of up to 20 off-the-shelf processor boards, configured around the IEEE-796 bus. Each board has a dual port memory, used for semaphore storage and distributed synchronization. Experiments using the automatic partitioner for five programs show a near-linear speedup, even for small problems, and a high utilization of the floating-point processor when the tasks size is large