Task-based Visualization Methods for Scalable Analysis of Large Data Sets
Final Report Abstract
In this project, the task-based parallelization of visualization algorithms, whole visualization pipelines and their performance on shared-memory and distributed architectures is investigated. The task-based approach differs to traditional parallelization approaches by its finegrained parallelism exposed by modelling data dependencies between small parts of the program, so called tasks. The advantage of this approach is that typical sources for inefficiency, like work-starvation and latencies, as they appear in traditional parallelization approaches like the prevalent bulk-synchronous model, can be circumvented or hidden. First, specific problems and visualization algorithms were investigated in isolation and corresponding, task-based algorithms were developed. Specifically, algorithms for visualizing scalar fields via iso-contours, scalar field topology via contour trees, vector fields via streamlines and rasterization and compositing algorithms were analyzed. Especially worth mentioning, an improved task-based algorithm for the computation of merge trees on distributed architectures was developed and shown to have better performance and scalability than previous approaches. This algorithm was further enhanced by an adaptive topological simplification approach to steer the size of the resulting output. Later, whole pipelines of task-based algorithms were investigated. For this, a framework was introduced that defines general structures for task-based algorithms to simplify the concatenation of succeeding pipeline modules. Furthermore, this framework defines communication patterns to allow for flexible data movement and manages the parallel execution and termination. With the help of this framework, pipelines consisting of the above-mentioned algorithms and a pipeline containing a direct volume renderer were analyzed with respect to their overall performance and the impact of interleaved execution of the task-based modules on the performance. In general, we could show that task-based visualization algorithms can be competitive to e. g. data parallel approaches. The investigation of concatenated task-based visualization algorithms yielded substantial improvements in comparison to implementations with no interleaving execution of modules in the observed pipelines, but the observed improvements were below expectations and appear capable of increase.
Publications
-
“Alternative Parameters for On-The-Fly Simplification of MergeTrees”. In: Eurographics Symposium on Parallel Graphics and Visualization. Hrsg. von Steffen Frey, Jian Huang und Filip Sadlo. The Eurographics Association, 2020. ISBN: 978-3-03868-107-6.
Kilian Werner & Christoph Garth
-
Unordered Task-Parallel Augmented Merge Tree Construction. IEEE Transactions on Visualization and Computer Graphics, 27(8), 3585-3596.
Werner, Kilian & Garth, Christoph
-
A Prototype for Pipeline-Composable Task-Based Visualization Algorithms. 2022 IEEE 12th Symposium on Large Data Analysis and Visualization (LDAV), 1-11. IEEE.
Petersen, Marvin; Werner, Kilian; Schnorr, Andrea; Kuhlen, Torsten Wolfgang & Garth, Christoph
-
In Situ Visualization for Computational Science: Background and Foundational Topics. Mathematics and Visualization, 1-8. Springer International Publishing.
Childs, Hank; Bennett, Janine C. & Garth, Christoph
-
“Towards Distributed Task-based Visualization and Data Analysis”. Diss. Technische Universität Kaiserslautern, 2022
Kilian Werner
-
“Extended Visual Programming for Complex Parallel Pipelines in ParaView”. In: Eurographics Symposium on Parallel Graphics and Visualization. Hrsg. von Roxana Bujack, David Pugmire und Guido Reina. The Eurographics Association, 2023. ISBN: 978-3-03868-215-8.
Marvin Petersen u. a.
