Lawrence Berkeley National Laboratory

One-dimensional decomposition of nonuniform workload arrays for optimal load balancing is investigated. The problem has been studied in the literature as "chains-on-chains partitioning" problem. Despite extensive research efforts, heuristics are still used in parallel computing community with the "hope" of good decompositions and the "myth" of exact algorithms being hard to implement and not runtime efficient. The main objective of this paper is to show that using exact algorithms instead of heuristics yields significant load balance improvements with negligible increase in preprocessing time. We provide detailed pseudocodes of our algorithms so that our results can be easily reproduced. We start with a review of literature on chains-on-chains partitioning problem. We propose improvements on these algorithms as well as efficient implementation tips. We also introduce novel algorithms, which are asymptotically and runtime efficient. We experimented with data sets from two different applications: Sparse matrix computations and Direct volume rendering. Experiments showed that the proposed algorithms are 100 times faster than a single sparse-matrix vector multiplication for 64-way decompositions on average. Experiments also verify that load balance can be significantly improved by using exact algorithms instead of heuristics. These two findings show that exact algorithms with efficient implementations discussed in this paper can effectively replace heuristics.