Using simulation and Markov modeling to select data flow threads

An algorithm is provided to give the set of all maximal sequential partitionings for a data-flow graph. Selecting an optimal partitioning from this set is incomputable. The authors explore two heuristics for selecting a partitioning: typical instance simulation and Markov analysis. In the two programs they tried, Markov heuristic selected the same partitionings as typical-instance simulation. Selecting a different partitioning had little impact on the execution time. However, using the worst-case maximal partitionings improved execution time by 33% and 18%, respectively, over the nonpartitioned programs. Using the best-case partitionings improved execution time by 39% and 20%, respectively.