Analysis of delays structure of interconnections in supercomputer by means of DBScan and divisive clustering algorithms

Abstract

In this paper we propose method for estimating and analysis measurements of delays in the computational cluster interconnection subsystem. Delays are combined into the set of pairs (source, destination). We have measurements of delays extracted by network_test2 utility from interconnections of following supercomputers: BlueGene/P, Lomonosov-1, Lomonosov-2 (Lomonosov MSU) and Jurope (Julich). We have clustered pairs of delays by DBscan and Divisive algorithms. Results of clusterisation revealed that DBScan is more accurate algorithm then divisive and allows to extract clusters, which correspond to the actual features in the supercomputer interconnections. Clusters gather near the same components of supercomputer network infrastructure. Gained clusters were visualized in 2-D by special tool, developed by authors.Introduction. Modern supercomputers are used to solve a wide range of problems, in particular: mathematical modeling and processing of large amounts of data. Supercomputers typically have an architecture of a computer cluster. Computations on supercomputers occur in parallel on many processing elements: cluster nodes equipped with multi-core processors, however, transfers between processors and cluster nodes can significantly slow down the parallel program. In order to minimize application performance losses, it is necessary to understand between which processors the delivery of messages occurs most quickly or vice versa slowly. It is necessary to optimally distribute the computations for the processors making up the supercomputer.