Распределенный алгоритм отображения распределенных многомерных данных на многомерный мультикомпьютер в системе фрагментированного программирования LuNA

Abstract

Distributed algorithm with local interactions Patch is presented in the paper. Patch is intended for data distribution and dynamic load balancing in the LuNA fragmented programming system. LuNA system is used for creation of parallel implementations of large-scale numerical models for distributed memory systems. Execution of a fragmented program is controlled by LuNA run-time system, which uses different data and computation distribution algorithms to enable efficient use of resources and minimize total execution time of the program. Patch algorithm, developed to be used in the LuNA system, enables distribution of multidimensional data meshes on a multidimensional lattice of computational nodes of a supercomputer. The algorithm uses mapping of data to multidimensional lattice of cells (coordinates), which in their turn are mapped to computational nodes. That mapping makes it possible to account for data dependencies and preserve data locality during dynamic load balancing. Patch algorithm was compared with another LuNA data distribution algorithm Rope, fragmented realisation of a real numerical problem was used for experiments. Experiments showed that Patch algorithm provides a general reduction in total computational volume and distances, as compared to Rope algorithm.