Design hierarchy-guided multilevel circuit partitioning

In this paper, we present a new multilevel circuit partitioning algorithm (dhml) which is guided by design hierarchy. In addition to flat netlist hypergraph, we use user design hierarchy as a hint for partitioning. This design hierarchy already has some implications on connectivity between logical blocks in the design. Using design hierarchy in partitioning is nontrivial since the hierarchical elements in design hierarchy do not necessarily have strong internal connectivity; hence, we need to determine whether it is preferable to break up or preserve the hierarchical elements. In order to identify and select the hierarchical elements with strong connectivity, their Rent exponents are used. Then, the selected hierarchical elements serve as effective clustering scopes during the multilevel coarsening phase. The scopes are dynamically updated (enlarged) while building up a clustering tree so that the clustering tree resembles the densely connected portions of the design hierarchy. We tested our algorithm on a set of large industrial designs in which the largest one has 1.8 million cells, 2.8 million nets, and 11 levels of hierarchy. By exploiting design hierarchy, our algorithm produces higher quality partitioning results than the state-of-the-art multilevel partitioner hMetis (Karypis et al. 1997). Furthermore, experimental results show that dhml yields significantly more stable solutions, which is helpful in practice to reduce the number of runs to obtain the best result.