In situ data-driven adaptive sampling for large-scale simulation data summarization
In situ data-driven adaptive sampling for large-scale simulation data summarization
复制标题
用于大规模模拟数据汇总的原位数据驱动自适应采样
DOI：
--
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
J. Ahrens
中科院分区：
文献类型：
--
作者：
Ayan Biswas;Soumya Dutta;Jesus Pulido;J. Ahrens
高性能计算的最新进展使科学家能够非常详细地模拟各种科学现象。然而，分析和可视化的输出数据，从这样的大规模模拟构成了巨大的挑战，由于其过大的大小和磁盘I/O瓶颈。这个问题的一个可行的解决方案是创建一个子采样数据集，它能够保留数据的重要信息，并且与原始数据相比，它的大小要小得多。创建用于生成这种智能子采样数据集的原位工作流程对于这种模拟是最重要的。在这项工作中，我们提出了一个信息驱动的数据采样技术，并将其与两个著名的采样方法进行比较，以证明所提出的方法的优越性。通过将其应用于Nyx宇宙学模拟，对所提出的方法的原位性能进行了评估。我们比较和对比这些不同的采样算法的性能，并提供一个整体的所有方法，使科学家可以选择适当的采样方案，根据他们的分析要求。
Recent advancements in high-performance computing have enabled scientists to model various scientific phenomena in great detail. However, the analysis and visualization of the output data from such large-scale simulations are posing significant challenges due to their excessive size and disk I/O bottlenecks. One viable solution to this problem is to create a sub-sampled dataset which is able to preserve the important information of the data and also is significantly smaller in size compared to the raw data. Creating an in situ workflow for generating such intelligently sub-sampled datasets is of prime importance for such simulations. In this work, we propose an information-driven data sampling technique and compare it with two well-known sampling methods to demonstrate the superiority of the proposed method. The in situ performance of the proposed method is evaluated by applying it to the Nyx Cosmology simulation. We compare and contrast the performance of these various sampling algorithms and provide a holistic view of all the methods so that the scientists can choose appropriate sampling schemes based on their analysis requirements.