SHF: Small: Developing a Highly Efficient and Accurate Approximation System for Warehouse-Scale Computers with the Sub-dataset Distribution Aware Approach

尽管当今仓库规模的计算机提供了巨大的数据处理能力，但从大数据集中获得即席查询答案仍然具有挑战性。为了解决这个问题，近年来出现了一种趋势，即利用近似计算在较小的原始数据样本上实现更快的执行，并在合理程度上牺牲结果的准确性。为了加速大数据查询，基于离线采样的方法和在线集群采样解决方案都已逐步部署到现实世界中。来自一流拉美裔博士学位授予机构的学生扩大了经验，并加强了计算机科学/工程课程活动，从而带来了教育方面的好处。关于仓库规模的计算机和大数据的在线跨院校本科选修课将有助于提供一种重新想象的学习体验，充分利用当今的技术，并由来自三所不同大学的学生提供广泛的富媒体学习材料。开发一个集成的硬件和软件、可扩展的近似系统存在重大困难。主要挑战是在给定的误差范围内将访问的数据的总大小及其关联I/O开销降至最低。现有流行的具有等概率解的整群抽样不能很好地处理许多遵循非均匀分布的真实世界的应用。本研究旨在通过研究子数据集分布感知的新方法来捕获子数据集分布，特别是非均匀类型的子数据集分布；采用非等概率聚类抽样来解决子数据集分布不均匀导致的采样效率低和方差大的问题；并考虑到采样过程的独特性质，以匹配计算机硬件特征，如SSD阵列，以充分释放其潜力。这项研究将确保未来的大数据近似系统能够实现高速的大数据分析，以革命性地改变人们与世界互动的方式；并通过高效和有效的数据处理提高经济影响的高生产率。

Despite the fact that today's warehouse-scale computers supply enormous data processing capacity, getting an ad-hoc query answer from a big dataset remains challenging. To attack the problem, recent years have seen one trend to exploit approximate computing to achieve faster execution on a much smaller sample of the original data by sacrificing result accuracy to a reasonable extent. Both offline based sampling approaches and online cluster sampling solutions have been gradually deployed in a real world to accelerate big data query. Educational benefits arise from broadening the experience of students from a top ranked Hispanic Ph.D. degree awarding institution and enhanced computer science/engineering curriculum activities. The online cross-institution undergraduate elective course about warehouse-scale computer and big data will be helpful in providing a re-imagined learning experience that makes optimum use of today's technologies supplemented by a broad range of media-rich study materials that students from three different universities. There are major difficulties in developing an integrated hardware and software, scalable approximation system.  The main challenge is to minimize the total size of accessed data and its associative I/O overhead subject to a given error bound. Existing popular cluster sampling with equal probability solutions do not deal well with many real-world applications following a non-uniform distribution. This research aims to tackle those challenges by investigating new sub-dataset distribution aware methods to capture sub-dataset distributions especially for non-uniform types, applying cluster sampling with unequal probability to address the inefficient sampling and large variance problem caused by non-uniform sub-dataset distribution, and taking into account the unique properties of sampling process to match with the computer hardware features, such as SSD arrays to unleash their full potential. The research will ensure future big data approximation system enables high velocity of big-data analytics to revolutionize the way that people interact with the world; and high productivity improvement of the economic impact through the efficient and effective data processing.