We study the problem of optimizing the mapping of LDPC codes on parallel machines to minimize the communication cost. To reduce the search space, the problem is solved in two stages: clustering, and cluster allocation. We propose a simplified clustering technique based on a modified min-cut algorithm that reduces the search complexity from O(n/sup 2/) to O(n). It was found that most of the locality is exploited by the clustering operation, which results in a 40-52% improvement in the total communication cost over random mapping. For large networks, cluster allocation is much more costly and results in only 1-8% additional improvement in unidirectional and bi-directional torus topologies. We compared the performance of two different approaches for cluster allocation. The first one is bused on min-cut algorithm, and the second one is based on a genetic algorithm. It was found that the min-cut based approach is better for small network sizes. For large network sizes with the number of clusters /spl ges/64, the genetic based approach becomes more attractive.
Discussion(0)
No comments yet. Be the first to comment.