GPU Accelerated Particle Visualization with Splotch

TL;DR

This paper discusses the redesign of the Splotch rendering algorithm to leverage GPU architectures, improving performance and scalability for large-scale particle datasets in astronomy and simulations.

Contribution

It introduces a GPU-accelerated version of Splotch using CUDA, with novel data organization schemes and a performance model to optimize rendering of large particle datasets.

Findings

Significant acceleration gains demonstrated on GPU implementations

Enhanced scalability for large datasets through optimized data management

Insights into parallelization challenges and solutions for GPU architectures

Abstract

Splotch is a rendering algorithm for exploration and visual discovery in particle-based datasets coming from astronomical observations or numerical simulations. The strengths of the approach are production of high quality imagery and support for very large-scale datasets through an effective mix of the OpenMP and MPI parallel programming paradigms. This article reports our experiences in re-designing Splotch for exploiting emerging HPC architectures nowadays increasingly populated with GPUs. A performance model is introduced for data transfers, computations and memory access, to guide our re-factoring of Splotch. A number of parallelization issues are discussed, in particular relating to race conditions and workload balancing, towards achieving optimal performances. Our implementation was accomplished by using the CUDA programming paradigm. Our strategy is founded on novel schemes achieving optimized data organisation and classification of particles. We deploy a reference simulation to present performance results on acceleration gains and scalability. We finally outline our vision for future work developments including possibilities for further optimisations and exploitation of emerging technologies.