[论文解读] Dark Sky Simulations Collaboration

Michael S. Warren, Warren, Michael S.|arXiv (Cornell University)|Jun 14, 2014

Advanced Database Systems and Queries被引用 2

一句话总结

Dark Sky Simulations合作项目发布了基于Titan超算上树状自适应方法的宇宙学N体模拟的早期公开数据，生成了超过55TB的数据，其中包括1.07万亿粒子的模拟。该模拟在1000倍质量范围内实现了质量函数和功率谱1%的自洽性，使暗物质与暗能量的精确实证研究成为可能，并为即将开展的大规模巡天提供应用支持。

ABSTRACT

The Dark Sky Simulations are an ongoing series of cosmological N-body simulations designed to provide a quantitative and accessible model of the evolution of the large-scale Universe. Cosmological simulations are the cornerstone of theoretical analysis of structure in the Universe from scales of kiloparsecs to gigaparsecs. Predictions from numerical models are critical to almost every aspect of the studies of dark matter and dark energy, due to the intrinsically non-linear gravitational evolution of matter. During the next few years, projects such as Pan-STARRS, the South Pole Telescope (SPT) and the Dark Energy Survey (DES) will measure the spatial distribution of large-scale structure in enormous volumes of space across billions of years of cosmic evolution. At the other extreme (sub-galactic and galactic scales from 100 parsecs to a megaparsec) understanding the distribution of dark matter within Milky Way type halos is necessary to interpret the results of Earth-based dark matter detection experiments. The revolutionary transformation of cosmology from a qualitative to a quantitative science has occurred over just the last twenty years. Driven by a diverse suite of observations, the parameters describing the large-scale Universe are now known to near 1% precision. Yet, the precise nature of dark matter and dark energy remain a mystery, and are unquestionably among the most important unsolved problems in physics. Advances in modeling must keep pace with observational advances if we are to understand the Universe which led to these observations. We have achieved superior performance on multiple generations of the fastest supercomputers in the world with our hashed oct-tree N-body code (<strong>HOT</strong>), spanning two decades and garnering multiple Gordon Bell Prizes for significant achievement in parallel processing. Using several new integrated and innovative algorithmic and computational science advances embodied in version 2 of the code (<strong>2HOT</strong>), combined with a unified data analysis effort based on the widely-adopted <strong>yt</strong> project, we propose a far-reaching set of scientific goals. We additionally aim to advance the state-of-the-art in domain decomposition and hierarchical tree-based computational techniques relevant to many simulation and data analysis problems. We will address a wide range of scientifically relevant tests of the standard cosmological model, including measurements of cluster abundance, void statistics, baryon acoustic oscillations, redshift-space distortions, velocity statistics and gravitational lensing. At small scales, we will test the abundance and central kinematics of the dwarf spheroidal galaxy satellites and related small-scale gravitational physics which determine the expected signal for dark matter detection experiments. Our simulations will produce an unprecedented suite of accurate and reliable halo, sub-halo and mock galaxy catalogs, which we will make publicly available.

研究动机与目标

• 为支持暗物质与暗能量的精确实证研究，提供一个公开可访问、高动态范围的宇宙学模拟套件。
• 克服分析模型在描述大尺度宇宙中非线性引力聚集方面的局限性。
• 通过模拟稀有且大质量的结构及其统计特性，为即将开展的大规模天空巡天提供精确预测。
• 为在互联网上分发的千万亿字节级模拟数据集，开发可扩展且高效的访问方法。
• 在计算完成后三个月内发布原始模拟数据，显著加速科学成果的传播。

提出的方法

• 在Titan超算上采用纯树状自适应N体方法（2HOT）进行引力力计算。
• 在(8h⁻¹Gpc)³体积内模拟了最多1.07 × 10¹²个粒子，使用了200,000个处理器。
• 使用yt进行数据分析与可视化，实现晕系照会、功率谱及光锥输出的提取。
• 实现了一种新型数据访问抽象机制，将网络映射为文件系统，支持远程内存映射访问与空间填充曲线索引。
• 生成了覆盖完整模拟体积1/10,000的光锥输出，以模拟类似巡天的视线观测。
• 通过不同盒子尺寸之间的内部一致性检查，以及与文献结果在1%–10%水平上的对比，验证了结果的可靠性。

实验结果

研究问题

• RQ1 纯树状N体方法是否足以在万亿粒子规模的宇宙学模拟中实现足够的精度与性能？
• RQ2 质量函数与功率谱在不同模拟盒子尺寸与粒子数下，其自洽性程度如何？
• RQ3 如何高效地通过互联网分发与访问千万亿字节级的模拟数据？
• RQ4 模拟生成的晕系照会与光锥是否能准确预测即将开展的大规模巡天的可观测量？
• RQ5 数据访问架构对大规模模拟数据在更广泛科研群体中的访问速度与可用性有何影响？

主要发现

• 在粒子质量范围超过1000倍的条件下，质量函数与质量功率谱实现了1%的自洽性，验证了模拟结果的内部一致性。
• 最大规模的模拟ds14a在(8h⁻¹Gpc)³体积内使用了1.07 × 10¹²个粒子，是迄今规模最大的宇宙学N体模拟之一。
• 与文献结果的对比显示，不同尺度下一致性在1%–10%之间，证实了模拟输出的准确性。
• 新型数据访问方法通过基于网络的文件系统抽象与空间填充曲线索引，实现了对单个34TB文件的远程高效访问。
• 团队在模拟完成后的三个月内发布了超过55TB原始模拟数据，显著加速了数据传播。
• 光锥数据集与相关晕系照会为预测巡天可观测量提供了独特资源，已通过Sunyaev-Zel’dovich星系团计数得到验证。