Performance debugging environments for parallel programs

Abstract

While writing a good parallel program is hard, writing one which is efficient is harder still. Though compiler technology is improving, we cannot expect it to turn all inefficient parallel programs into efficient ones. This is due to the fact that parallel algorithms can fail in many more dramatic ways than sequential algorithms running on conventional hardware. Consequently, a powerful performance debugging system must be an integral part of a parallel program development environment. Most of the existing parallel program debugging approaches have been ad-hoc and tailored to specific machines. In this work we develop a novel approach for developing portable and programmable performance debugging environments, which addresses a number of important issues including instrumentation, performance visualization and scalability studies. Central to the approach is the use of the program dependence graph and the treatment of the program performance data as a historical database. We demonstrate with the help of examples that a systematic application of program views can help in debugging the performance of programs. We also develop a method for reducing the perturbations introduced in a program execution, due to source level instrumentation, using event and time based analysis.

Cite this paper

@article{sarukkai1993performance,
  title = {{Performance debugging environments for parallel programs}},
  author = {Sarukkai, Sekhar R.},
  year = {1993}
}

Made from the index record. Check names, capitalisation and page numbers against the paper. Need another style? AI for Research has over 10,000.

Related papers