Optimization of mapping non-uniform interacting processes onto the architecture of computing cluster

The project aims to optimize the projection of non-uniform interacting MPI processes on the architecture of computing cluster. In the course of the project, the method for mapping non-uniform interacting parallel processes of a computing application, which uses MPI, onto cluster architecture has been developed. The method is intended to reduce delays during synchronization by assigning the most intensively interacting processes to computing cores with the fastest interconnect. The method uses the representation of the computational task and cluster architecture in the form of a weighted graph. The heuristic algorithm, which gives the result of mapping process numbers onto the numbers of cluster computing cores in an acceptable time, has been developed. Acceleration of calculations by 16-20% has been obtained by optimizing the mapping for the tests of 300-4800 processes and shown with using a well-scalable computing package CORAL QBOX as an example.