| Abstract |
|---|
| Distributed memory machines are still not broadly accepted although they provide large computing power. One of the reason is the costly development process for a specific parallel algorithm on a specific distributed memory machine. Due to the complicated runtime behavior caused by communication overhead and load imbalance finding an efficient parallel version of an algorithm often takes a lot of time. There is request for efficient and accurate runtime functions which easily predict the performance of the algorithms. The contribution of this paper is to show that runtime functions predicting the execution time of the communication operations necessary for any performance prediction mechanism can be generated by means of the genetic programming paradigm. The runtime functions generated dominate the runtime functions presented in literature, till today. |
| Contact |
| Laura Heinrich-Litan Institut fuer Informatik,Universitaet Halle,Kurt-Mothes-Str. 1,D-06099 Halle (Saale),GERMANY laura@informatik.uni-halle.de |