A Course-Grid Pre-Processing based Load Balancing Scheme for Parallel Line Integral Convolution

Line Integral Convolution (LIC) has proven to be a very powerful visualization technique of vector data sets such as those obtained from numerical simulations. Despite its usefulness, this technique is notoriously computational intensive and is increasingly challenging as the amount of data continuously grows due to the advances in both computer hardware and software technology. Parallel processing is considered an effective approach for handling such large data sets, however it can suffer from workload imbalance depending on the distribution and orientation of the vectors elements in a given vector field. Specifically, elements on the boundary as well as close to the singularities (sources or sinks) have the potential for causing workload imbalance. In order to minimize this problem, we propose a load balancing scheme which adjusts the data partitioning based on the estimated workload obtained by coarsegrid LIC pre-processing. The partial results obtained from this pre-processing can be fully utilized in the next stages thus minimizing the pre-processing overhead. We obtained encouraging results showing the suitability of this load balancing scheme for parallel LIC which greatly contributes to the improvement of the visualization process of large vector data.