Pairwise Grouping of Adjacent Nodes

Abstract

In the original Pairwise Grouping of Adjacent Nodes (PGAN) technique for joint code and data minimization, a cluster hierarchy is constructed by clustering exactly two adjacent vertices at each step [Bhat93]. At each clustering step, a pair of adjacent actors is chosen that maximizes ρ({A, B}), the repetition count of the adjacent pair, over all clusterable adjacent pairs {A, B}. Recall from Section 4.3 that ρ(Z) can be viewed as the number of times a minimal periodic schedule for the subset of actors Z is invoked in the given SDF graph, and thus, we see that the PGAN technique repeatedly clusters adjacent pairs whose associated subgraphs are invoked most frequently in a valid schedule.