Interaction Patterns Detection in PVM Programs to Support Simulation

In this paper we propose a solution, based on static analysis and statistical techniques, for the problem of the determination of the possible interaction patterns among the processes of a PVM message passing program, when these present a data-dependent behavior. A prototypic implementation of this technique has been coupled with MPSS, a tool for simulation and performance prediction of PVM message passing programs, in order to overcome its unability to deal with non-deterministic and data dependent programs, and thus to provide for a complete program characterization in terms of idle-, cpu-, communication and synchronization time, for irregular and data-dependent programs.