A decomposition approach for balancing large-scale acyclic data flow graphs

An efficient decomposition technique that provides a more systematic approach in solving the optimal buffer assignment problem of an acyclic data-flow graph (ADFG) with a large number of computational nodes is presented. The buffer assignment problem is formulated as an integer linear optimization problem that can be solved in pseudopolynomial time. However, if the size of an ADFG increases, then integer linear constraint equations may grow exponentially, making the optimization problem more intractable. The decomposition approach utilizes the critical path concept to decompose a directed ADFG into a set of connected subgraphs, and the integer linear optimization technique can be used to solve the buffer assignment problem in each subgraph. Thus, a large-scale integer linear optimization problem is divided into a number of smaller-scale subproblems, each of which can be easily solved in pseudopolynomial time. Examples are given to illustrate the proposed decomposition technique