A Task Merging Technique for Parallelization of Modelica Models

This paper presents improvements on techniques of merging tasks in task graphs generated in the ModPar automatic parallelization module of the OpenModelica compiler. Automatic parallelization is performed on Modelica models by building data dependency graphs called task graphs from the model equations. To handle large task graphs with fine granularity, i.e. low ratio of execution and communication cost, the tasks are merged. This is done by using a graph rewrite system(GRS), which is a set of graph transformation rules applied on the task graph. In this paper we have solved the confluence problem of the task merging system by giving priorities to the merge rules. A GRS is confluent if the application order of the graph transformations does not matter, i.e. the same result is gained regardless of application order. We also present a Modelica model suited for automatic parallelization and show results on this using the ModPar module in the OpenModelica compiler.