Increasing the performance of cluster systems (HPC) while taking into account network latency

This report is devoted to the analysis of ways to increase the performance of supercomputers without upgrading their hardware. By studying larger-scale routing, we found that network latency can be used as a metric function. The principle of minimizing latency when assigning new nodes for parallelization leads to a significant increase in performance, up to 20%. The dynamics of growth of power grid delay depending on the size of the administrative package has been studied.