Methods for Automatic Detection of the Sufficient Number of Iterations for Delay Measurements in Computer Cluster Interconnections

An MPI benchmark is a specialized software used for testing the network environment of a supercomputer cluster. When running such programs in a selected test mode, the user sets several parameters. This setting depends on the type and volume of the network load provided by the selected test mode. Benchmarks usually make delay measurements while data are transferred with MPI functions. The delay values for MPI messages may vary over time. Therefore, in the set of parameters, there is almost always a parameter responsible for the number of delay measurements. This parameter is crucial for the information content of collected data. We provide some criteria that allow users to evaluate the information content of collected data. Algorithms based on these criteria strive to determine the number of measurements at runtime. The benchmarking process stops if collected data meet the selected criterion, allowing users to get more informative data and reduce benchmarking time. The features of the proposed algorithms have been investigated at the Shared Research Facilities “High-Performance Computing and Big Data” (“CKP Informatics”) of the Federal Research Center “Computer Science and Control” of the Russian Academy of Sciences (FRC CSC RAS, Moscow). The algorithms reduce the number of iterations by half with the following specified qualities: 0.75 for the scalar-based algorithm and 0.81 for the Fourier-based algorithm.