On the Performance of Line Integral Convolution in a Distributed-Memory Parallel Setting

Abstract

Line integral convolution (LIC) is a powerful tool for visualizing vector fields by combining particle advection with image convolution. Practical usage of LIC is limited by its computational expense, requiring many calculations for every cell in the mesh. Fortunately, computation of LIC can be accelerated through parallelization. In this thesis we evaluate whether LIC parallelizes better over distributed systems than comparable particle advection algorithms. We do this by harnessing the VisIt Parallel Integral Curve System for the generation of LIC convolution kernels. We also contribute an extension to LIC which reduces dependency on input data. We look at how the algorithm compares to other advection techniques with respect to performance and load balancing. We evaluate the performance of LIC with PICS across 36 different test configurations with three metrics. We find a 2x performance improvement and an up to 8x load balancing improvement for LIC over traditional parallel streamlines.