With the scale of supercomputing systems and applications growing fast, the difficulty of developing performance efficient applications also grows rapidly. The reason for this is an extensive number of factors that potentially influence the application performance. Hardware and software specifics of the supercomputer, peculiarities of the application, interference of jobs running simultaneously - everything needs to be taken into account when trying to achieve high performance. With supercomputers constantly evolving, all these specifics become more and more complicated. This indicates the demand for a specific tool that would allow seeing where and, what is more important, why does the performance loss happen. In this paper we give an overview of the developed toolkit and discuss in detail one of the approaches aimed at studying the application behavior during the job run. This approach studies the dynamic characteristics of jobs that are gathered by monitoring tools. Its aim is to provide system administrators and users with overall job characteristics in order to get both overall and detailed analysis of every separate job run. This approach and the generated detailed report have been named "Job Digest".

Keywords: supercomputer, performance, efficiency study, monitoring, parallel computing, dynamic job characteristics, high performance computing

• PDF (1.24MB)



• PDF. zip (1.04MB)