Automated modeling of message-passing programs

We present a system for automated modeling of message-passing programs. Its models preserve the parallel program´s structure, especially the syntactic boundaries surrounding communication calls. Our grammar-driven approach uses the program´s parse trees to derive a regular expression that describes all possible execution traces at the chosen level of modeling; that expression is used for automatic extraction of timing information from traces of scaled-down runs. We consider “intelligent regression” techniques for discovering the numerical attributes of our models: run times of sequential blocks; lengths and destinations of messages; and loop bounds. Regression produces formulae expressing these attributes in terms of problem and system sizes. The model is then used for predicting the performance of large-scale runs. We illustrate our approach with a program that simultaneously solves multiple tridiagonal linear systems an the iPSC/860