NHR PerfLab Seminar: Performance Measurements in HPC

Symbolic picture for the article. The link opens the image in a large view.

Speaker: Dr. Christian Iwainsky (TU Darmstadt)

Title: Performance Measurements in HPC: Techniques, Side Effects, and Use

Abstract:

Optimizing and tailoring an HPC code to a given system and problem requires a good understanding of code’s hardware utilization. Performance measurements are the essential means for gaining such critical insights. Here, different metrics are used to assess whether a given code segment makes a good use of given hardware resources. However, mastering available measurement systems requires understanding of the measurement techniques and data processing systems used. In particular, control of measurement overhead and a balance of perturbance with measurement expressiveness is critical for useful insight.

In this talk we deep-dive the technology concepts of current performance analysis tools to gain a critical understanding of their capabilities and limitations, and the overheads and perturbation involved. This insight enables one to utilize these tools to their fullest potential in order to write efficient code and optimize HPC programs.

Slides

Speaker bio:

Dr. Christian Iwainsky is a senior research scientist at TU Darmstadt and HPC expert at the Competence Center for High Performance Computing in Hessen. He has been successfully involved in the performance analysis and optimization of several large-scale HPC codes from various domains. Currently, he is applying this expertise to OpenFOAM in an EU-funded research project and at the national high performance computing center for computational engineering sciences. His research interests center on parallel programming and related software engineering aspects with a focus in leveraging compiler technology in performance analysis.