Seminar

Modern supercomputers provide a lot of useful information about the applications running on them: data on the structure, performance or communication profile of applications; names of the used application software packages, libraries and compilers; detailed information on the launch of jobs, etc. The volume of collected information is growing, and it is almost impossible to process it manually. Therefore, the problem of developing data mining methods are becoming increasingly relevant which will allow administrators to more fully, accurately and quickly evaluate the work of a supercomputer based on the specified information, as well as to identify and eliminate problems that lead to a decrease in the efficiency of supercomputers. One of the areas for such analysis is the problem of finding similar applications. Having information about the similarity of various applications, it is possible not only to study new jobs using the previously obtained results of the analysis of similar, already studied applications, but also to group jobs or predict their behavior, which will significantly facilitate the process of studying the efficiency of applications for both users and administrators of supercomputers. This research presents two approaches to solving the problem of finding similar supercomputer applications, and also proposes algorithms for studying the supercomputer job flow based on the proposed approaches, which allow identifying the software package usages, job clustering, and predicting the quality assessment of supercomputer resources usage.

Сonnecting to MTS Link.
Information on the seminar and the link to connect are available at Indico.