Implementation of codebook generation algorithm on a campus PC Grid

Abstract

PC Grid system is a promising platform to achieve a high performance computing with less construction cost and less power consumption. Parameter sweep applications (PSA) can be efficiently carried out on such a parallel system because they include few communications, but some applications are inappropriate for the parallel system due to a large communication cost. We focus on PNN (Pairwise Nearest Neighbor) codebook generation algorithm for VQ (Vector Quantization) compression algorithm as an example of hierarchical clustering methods and proposed a parallel version of the PNN algorithm suitable for the PC Grid system with distributed memory, called multi-step parallel PNN. However, the computational complexity of the second half of this algorithm increases as the number of worker computers grows, so we apply a tree-distribution implementation to the second half in order to cope with this problem. We call this approach parallel PNN. We confirm the effectiveness of the reduced parallel PNN by the evaluation of the computational complexity of the algorithm and the experiment executed on a campus-wide PC Grid with over 60 worker computers.