Research into Changes in the Level of System Noise in High-Performance Clusters
Daria Domracheva, Konstantin Stefanov · Lobachevskii Journal of Mathematics · 2022
Abstract This study aims to investigate system noise in high-performance systems. Noise in high-performance clusters can negatively impact application performance, increasing application runtime. To detect noise and analyze its behavior, it is proposed to use a method based on the work of collective MPI operations, and the noise can be analyzed both at one node and at several. The results obtained in this way are compared with the results of the benchmark utility Netgauge. Measurement data can be clustered by duration using algorithms such as BIRCH and DBSCAN, which makes it possible to highlight certain parameters inherent in a particular noise cluster. It is also shown that some of the clusters obtained as a result of the analysis can be associated with various system parameters that can cause noise.