A new clustering algorithm for large communication delays

We are interested in studying the clustering problem in the case of large communication delays for an unbounded number of processors. It has been extensively studied as a basic step for obtaining efficient algorithms for scheduling the tasks of a parallel program. We present a new approach for solving this problem based on a recursive decomposition of the precedence task graph. We first establish a general result on a class of clustering algorithms with specific properties called convex clustering that can theoretically be proved to be within a factor of 2 from the optimal. Then, we propose an algorithm for building convex clustering. The idea is to determine two sets of independent tasks as large as possible that can lie executed simultaneously without communications and to apply it recursively on each set. This approach is assessed by some simulations run on some families of structured task graphs.