数千上万核上模拟结果的并行可视化

Parallel Visualization of the Simulations on Tens of Thousands of Cores

在科学与工程计算中,在数千上万核上,模拟包含上亿网格单元的模型通常输出TB量级的时变数据集。这些数据集包含多个时刻的数据,每个时刻的数据分布存储在数千个文件,单时刻的数据量达到几GB甚至几十GB。为了并行地可视化这种时变数据集,本文设计了基于网格片的层次化数据结构,基于该数据结构改进和优化了并行可视化流程和数据通信算法。在集成到可视化软件后,实现了两个应用的模拟结果可视化。实测数据表明,对于单时刻5GB和32GB的数据,在数十上百个处理器核上,从数据读入到获得体绘制结果的时间分别为19秒和80秒,交互操作为2秒到10秒;三维面绘制在数秒钟内,切片分析在2秒以内。

In scientific and engineering computing, some datasets generated by the simulations on tens of thousands of cores are measured in terabytes or even more. These datasets contain the data from multiple timesteps. The amount of each timestep dataset distributed to thousands of files is gigabytes or even tens of gigabytes. The hierarchical data structure based on the patch is designed to support the visualization of such scale dataset. And more, the parallel visualization pipeline and data communication algorithms are improved and optimized. Finally, two datasets coming from the code run on tens of thousands of cores are visualized. The measured result show that for the 5GB and 32GB data at single moments, on tens of thousands of processor cores, the time between the data imput and the result output is 19s and 80s respectively, and the interactive operation is between 2s and 10s; for the 3D rendering it is within several seconds, and the slicing is within 2s.