Dimemas: predict parallel performance using a single cpu machine

Dimemas is a performance analysis tool for message-passing programs. It enables the user to develop and tune parallel applications on a workstation, while providing an accurate prediction of their performance on the parallel target machine. The Dimemas simulator reconstructs the time behavior of a parallel application on a machine modeled by a set of performance parameters. Thus, performance experiments can be done easily. The supported target architecture classes include networks of workstations, single and clustered SMPs, distributed memory parallel computers, and even heterogeneous systems.

Several message-passing libraries are supported, including PVM, MPI and PARMACS. For communication, a linear performance model is used, but some non-linear effects such as network conflicts are taken into account. The simulator allows specifying different task to node mappings.

Dimemas generates trace files that are suitable for two performance analysis tools: Paraver and Vampir. Paraver and Vampir enable the user to conveniently examine any performance problems indicated by a simulator run.

The analysis module performs critical path analysis reporting the total CPU usage of different code blocks, as well as their importance for the program execution time. Based on a statistical evaluation of synthetically perturbed traces and architectural parameters, the importance of different performance parameters and the benefits of particular code optimizations can be analyzed.

Dimemas main goals can be sumarized as:

  • Utilization
  • Development and tunning
  • Benchmarking
  • Training
  • Features
  • Predict behaviour of target architecture
  • Forecast effects of code optimization
  • What if analysis
  • Usage
  • Integrated with visualization tool
  • Extensive application characterization
  • Squeezing the information that a single instrumented run can provide