Clustering of results collected by benchmarking application of multiprocessor system interconnect

Abstract

A number of processing units in a supercomputer constantly grows leading to increased complexity of a communication network. In order to minimize performance lose because  of communication delays and in order to ease understanding behaviour of communication environment in supercomputer some approaches are used.  One  collect message passing statistics where other concentrates on post-run analysis of collected information. Our approach is based on measuring of delays which arise during the messages passing. Benchmarking program produces a huge amount of data that can not be analysed manually. Problems of compression, visualization and automated analysis such data are solved with use of clustering methods. Results of analysis are usable for developing and optimizing code of parallel programs and to detect communication environment anomalies. In this paper we discuss the advantages and disadvantages of algorithms: division, agglomerative, local rank  clustering application on data received from the Lomonosov, SKIF-MSU, Bluegene/P MSU supercomputers.