Automatable & scalable late cycle performance analysis

  • Authors:
  • Florian Mangold;Moritz Hammer;Harald Roelle

  • Affiliations:
  • LMU PST, Munich, Germany;LMU PST, Munich, Germany;Siemens AG CT SE 1, Munich, Germany

  • Venue:
  • Proceedings of the first joint WOSP/SIPEW international conference on Performance engineering
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

Performance analysis of large, concurrent systems is a difficult problem that can hardly be approached with classical profiling. Performance issues might be caused by the interaction of modules and hardware components, making it difficult to find exact causes by considering single modules. By slowing down single modules artificially, dependencies of modules can be detected. Employing statistical means, such dependencies are detected in the covariance of runtime changes. We propose a way to detect the most meaningful dependencies in large-scale systems, allowing arbitrary scaling with respect to the granularity considered.