Table 3. Comparison of Bisection Solution Quality, Using Actual Module Areas. 4. Conclusions and Future Work

Table 2. Comparison of bisection solution quality, using unit module areas. each run consists of 1000 passes of FM. LSMC outperforms the previous best reported results in the literature by 6.4 for unit area and averages 21.4 improvement o ver FMx50. With actual module areas, LSMC is 1 better than previous results, and averages 36 improvement o ver FMx50. We h a ve presented an experimental study of several local perturbation strategies kick m o ves for the Large-Step Markov Chain LSMC heuristic applied to VLSI netlist bisection. Our experimental results show t h a t t h e LSMC heuristic consistently outperforms the standard random multistart strategy, g i v en the same computational resources. The quality of our results compare favorably to existing published results. We are currently pursuing the following extensions to our work: 99 showed that for the TSP, the eecacy of the kick moves depends on the underlying greedy descent engine. This may be the case for netlist partitioning as well. We are currently studying whether implementation details such a s l o o k ahead 111 or the gain bucket tie-breaking scheme used 66 will signiicantly aaect the choice of kick m o ve. Problem reduction techniques can signiicantly improve the performance of iterative improvement partitioning algorithms 77. In such hierarchical approaches, it is possible to use LSMC in place of the standard FM algorithm to improve solution quality, at the cost of additional runtime. For example, in a 2-phase FM approach 33, the calls to the FM algorithm can be replaced with calls to LSMC. We are currently experimenting with this approach; preliminary experiments show that this is quite promising. The LSMC heuristic can be straightforwardly generalized to k-way partitioning each of the kick m o ves described in Section 2 has an obvious k-way generalization. In preliminary experiments, k-way LSMC with clustering kick m o ve is promising, yielding better results than equivalent computational eeorts with a m ultistart k-way F M. Finally, , 9 has shown that more elaborate variants of LSMC, e.g., using non-zero temperature schedules and population-based optimization can be eeective o n t h e TSP. Similar improvements may be found for netlist partitioning.. cut; after performing greedy descent in the reduced netlist, Nu is restored. The combination of these operations constitutes the kick m o ve. In our implementation MoveSize, the percentage of uncut nets removed, is …