Program implementation based on macro data flow paradigm with RDMA communication support

摘要

In this paper, we discuss execution methodology for parallel fine grain programs based on the macro data flow paradigm applied to a FDTD program which is an example of fine grain parallel application based on regular computations executed in an irregular domain. Parallel applications are executed in a MIMD system with message passing implemented with RDMA facility based on rotating buffers control infrastructure. It is shown that such execution model for fine grain parallel applications can facilitate control and synchronization of resources involved in computations and communication. Execution based on macro data flow paradigm reduces synchronization overhead which can not be avoided in message passing communication. This is achieved at a cost of processor time spent on monitoring of program macro node states since we use here a traditional von Neuman system with architectural model unsupported for macro data flow execution. To achieve the best speedup, assignment of macro nodes to physical processors is proceeded by static analysis of program code and optimal decisions as regards node definition/allocation have to be taken.