A taskgraph clustering algorithm based on an attraction metric between tasks

Abstract

Task granularity is a critical issue in the partitioning of a parallel program. To adjust automatically the task granularity for the target system, the solution used for the grain-size problem is the bottom-up approach: first the program is partitioned into elementary operations and mathematical functions and then several operations are combined into larger tasks. The conglomeration of tasks is controlled by attraction values: the attraction between two tasks is proportional to the benefit of aggregating the two tasks. The clustering heuristic is embedded in the definition of the attraction between tasks as only tasks with an attraction value higher than a certain threshold are conglomerated. It is assumed that the task graph structure and the task lengths are known at compile time. This information is used by the clustering algorithm. The algorithm defines an attraction between two tasks and then conglomerates the tasks for which the attraction is larger than a given threshold.<>