An MILP Based Algorithm for Automatic System Level Synthesis

Automatic Synthesis of large realtime systems using an array of oo-the-shelf general purpose processors, digital signal processors and Field Programmable Gate Arrays is increasingly gaining importance. Each of the subproblems in such a synthesis process, namely, pipelining, resource type selection, resource allocation, resource assignment, resource scheduling and resource re-connguration are complex and are conventionally solved using a mixture of exact and heuristic techniques (often applied in a phased manner) with little guarantee on the optimality of the solution. In contrast, this paper formulates the pipelining, selection, allocation, scheduling and recon-guration problems as a single mixed integer linear programming (MILP) problem to ensure optimality. We show that for many practical size problems, our formulation can produce optimal solution in reasonable time. We believe that the uniqueness of our technique stems not only from the optimality of the solutions it can generate but also from the fact that the problem we have modeled is a generalization of several synthesis problems. Compared to the conventional MILP models for scheduling, the number of integer variables needed in our model are much less and is independent of the total cycle time (which can be very high in system level synthesis). Our model can handle time constrained, resource constrained as well as time and resource constrained synthesis problems.