HeROD flavoured oct-trees: scientific computation with a multicomputer persistent object store

Abstract

Using a persistent multicomputer object store should greatly simplify the writing of distributed memory parallel programs operating on irregular, object-structured data, by removing from the programmer the burden of managing data referencing, distribution and coherency. In this report we explore the development of such a program using the HeROD persistent multicomputer object store, in a computational science application. The computational code is part of the tree-code algorithm for the N-body problem. A number of solutions to the difficulties of combining flat transactions and cooperating parallel processes are explored, in both the application programming domain and the persistent store design. Actual performance measures of the implementation on a 128-processor multicomputer are reported, with scalability comparisons for a range of processor configurations (exploring sensitivity to the client-server balance) and program implementation strategies, particularly by varying transaction size.