Abstract: In this paper, we report the implementation and measured performance of global simulation of planetary rings on Sunway TaihuLight. The basic algorithm is the Barnes-Hut tree, but we have made a number of changes to achieve good performance for extremely large simulations on machines with extremely large number of cores. The measured performance is around 35% of the theoretical peak. The main limitation comes from the performance of the interaction calculation kernel itself, which is s currently around 50%.

Abstract: Modelling and simulation of bacterial biofilms is a computationally expensive process necessitating use of parallel computing. Fluid dynamics and advection-consumption models can be decoupled and solved to handle the fluid-solute-bacterial interactions. Data exchange between the two processes add up to the communication overheads. The heterogenous distribution of bacteria within the simulation domain further leads to non-uniform load distribution in the parallel system. We study the effect of load imbalance and communication overheads on the overall performance of simulation at different stages of biofilm growth. We develop a model to optimize the parallelization procedure for computing the growth dynamics of bacterial biofilms.

Abstract: Clustering algorithms are essential for many big data applica- tions involving point-based data, e.g. user generated social media data from platforms such as Twitter. One of the most common approaches for clustering is DBSCAN. However, DBSCAN has numerous limitations. The algorithm itself is based on traversing the whole dataset and identi- fying the neighbours around each point. This approach is not suitable when data is created and streamed in real-time however. Instead a more dynamic approach is required. This paper presents a new approach, RT- DBSCAN, that supports real-time clustering of data based on continuous cluster checkpointing. This approach overcomes many of the issues of existing clustering algorithms such as DBSCAN. The platform is real- ised using Apache Spark running over large-scale Cloud resources and container based technologies to support scaling. We benchmark the work using streamed social media content (Twitter) and show the advant- ages in performance and flexibility of RT-DBSCAN over other clustering approaches.

Abstract: X-ray imaging allows biologists to retrieve the atomic arrangement of proteins and doctors the capability to view broken bones in full detail. In this context, ptychography has risen as a reference imaging technique. It provides resolutions of one billionth of a meter, macroscopic field of view, or the capability to retrieve chemical or magnetic contrast, among other features. The goal is to reconstruct a 2D visualization of a sample from a collection of diffraction patterns generated from the interaction of a light source with the sample. The data collected is typically two orders of magnitude bigger than the final image reconstructed, so high performance solutions are normally desired. One of the latest advances in ptychography imaging is the development of Ptycho-ADMM, a new ptychography reconstruction algorithm based on the Alternating Direction Method of Multipliers (ADMM). Ptycho-ADMM provides faster convergence speed and better quality reconstructions, all while being more resilient to noise in comparison with state-of-the-art methods. The downside of Ptycho-ADMM is that it requires additional computation and a larger memory footprint compared to simpler solutions. In this paper we tackle the computational requirements of Ptycho-ADMM, and design the first high performance multi-GPU solution of the method. We analyze and exploit the parallelism of Ptycho-ADMM to make use of multiple GPU devices. The proposed implementation achieves reconstruction times comparable to other GPU-accelerated high performance solutions, while providing the enhanced reconstruction quality of the Ptycho-ADMM method.