Navigation and service

Tutorial "Large-Scale Application Performance Analysis using Scalasca"

27 June 2007
Central Institute for Applied Mathematics (ZAM) of Forschungszentrum Jülich

As the size of today's capability computing systems and the number of processors used by individual applications is continuously raised, achieving satisfactory performance and scalability becomes more and more difficult. To improve the efficiency and, thus, to expand the potential of large-scale applications, powerful tools are needed that collect relevant data on code performance and identify causes of performance problems. However, traditional tools often cease to work in a satisfactory manner at a scale of thousands of processors.

In this full-day tutorial, we will provide a comprehensive introduction to the newly developed Scalasca toolset, which specifically addresses these scalability requirements. Being the successor of the well-known KOJAK automatic performance analysis tool, it can automatically detect inefficient program behavior and highlight opportunities for performance improvement. It has already demonstrated its applicability on various large-scale supercomputers such as IBM BlueGene/L and Cray XT3/4 up to 20,000 processes. Currently, Scalasca supports applications based on the Message Passing Interface (MPI) Version 1. For more information on Scalasca, please see

The scope of the tutorial covers all important steps of the performance analysis cycle, including:

  • Instrumentation
  • Runtime summarization

    • Call-graph execution time
    • MPI message statistics
    • Hardware counter profile
  • Event trace collection
  • Automatic trace analysis

    • Quantification of wait states
    • Identification of potential message-ordering inefficiencies
  • Performance algebra utilities

All steps will be demonstrated with real-world examples. In addition, we will have hands-on exercises where attendees can do their own measurements and analyses using the Scalasca tools with provided example codes. We expect to use JUBL (BlueGene/L) and JUMP (SP2 cluster) as target platforms: participants should have accounts and basic familiarity with running applications on one or both systems.

The tutorial is free of charge and will be held at the Central Institute for Applied Mathematics (ZAM) of Forschungszentrum Jülich on Wednesday, June 27th 2007 between 9:30 and 16:30. Participation is limited: to register, please send an e-mail to Markus Geimer by June, 18th 2007. If you have a time conflict, but would like to attend, please send us a note so that we can inform you when we will repeat this tutorial.

This tutorial is organized by the Helmholtz-University Young Investigators Group "Performance Analysis of Parallel Programs".

Markus Geimer
Brian Wylie