Adaptation of Parallel Computational Algorithm to Supercomputer Architecture by Example of Plasma PIC Simulation

In order to increase the efficiency of supercomputer engagement and to accelerate 3D turbulent plasma simulation it is necessary to provide an on-the-fly adaptation of the algorithm to the computer architecture. A number of 3D computational experiments were conducted using the two-stage Euler-Lagrangian decomposition. In order to adopt the program to computational nodes provided for the computation two new approaches were proposed. First, measure the time of collective and peer-to-peer MPI operations in a small number of test runs and then make and estimation of the operation time dependency on the number of processors and finally figure out the optimal combination of Eulerian and Lagrangian decomposition. Second, analyze the computational nodes provided for the program execution and reorder the MPI processes considering their position within the nodes.