Performance Analysis and Optimization

Performance analysis and optimisation of a parallel application can be a difficult task, and often it is not clear where to start. For instance, we might ask if the chosen communication pattern is efficient enough? Does my application suffer from load imbalance? Does the way computational work is divided is a problem? Or does something else impact performance? During this session you will learn basic ideas about performance engineering. You will get acquainted with different ways and tools for performance measurement. During hands-on you will try to collect, analyse and interpret measurements of a prepared benchmark. 

Ilya Zhukov is an HPC applications analyst at Jülich Supercomputing Centre (JSC) at Jülich Research Centre, based in Germany. His main focus is to help researches and developers of HPC applications to gain in-depth understanding into the performance of HPC applications, identify limiting factors, and provide feedback to improve the efficiency of HPC applications on current and potentially future systems. Besides he teaches on topics in performance analysis and usage of performance analysis tools.

Preparation

• Please check and fulfill prerequisites prior to the hands-on sessions

Course content

• Introduction to Performance Engineering
  
• Single node optimisation
  
• BT-MZ reference
  
• Score-P measurement system
  
• Cube visualisation browser
  
• Score-P scoring and filtering
  
• Automatic trace analysis with Scalasca
• Vampir, TAU, Extra-P, Darshan
• Optional: Case studies (please download profiles here)
  
• VI-HPS Overview and POP
  

Prepared measurements

• Bridges
  
• Comet