|
Abstract : |
Abstract. In this paper we explain our strategy for parallelizing a wavelet based compression routine. The compression routine is designed to work within a earthquake simulation model where huge amount of data is written and read from disks. As the compression routine makes up for a significant part of the computation it is very important that it parallelize well. We have implemented the parallel version in OpenMP. This makes the implementation easy, but it hides a fundamental problem. The optimal data layout on a distributed memory system changes for the di#erent parts of the algorithm. Thus either data has to be redistributed or the processors have to do remote data access. Our experiments show that the remote data access work pretty well on the Origin system thus for the number of CPU less or equal to 32, no significant slow down, due to remote data access is seen., |
