SHF: Small: Collaborative Research: ALETHEIA: A Framework for Automatic Detection/Correction of Corruptions in Extreme Scale Scientific Executions

信任科学应用需要保证计算结果的有效性。不幸的是，许多科学计算的例子导致了不正确的结果，有时会带来灾难性的后果。目前已知的验证技术只涵盖了数值模拟和数据分析应用程序在执行过程中可能遭受的损坏的一小部分。随着科学过程的规模和复杂性的增长，其组成步骤的可靠性和有效性越来越难以确定。在存在潜在数据损坏的情况下评估有效性是一个严重且未得到充分认识的问题。损坏可能发生在从硬件到应用程序的所有计算级别上。这些腐败的一个重要方面是，在它们被发现之前，所有的执行都有被悄无声息地破坏的风险。在一些有记录的案例中，从发现损坏到通知用户已经经过了几个月。与此同时，可能会有大量的执行被破坏，从而导致错误的结论。事后可能很难检查执行是否确实存在腐败，因此，即使腐败不会导致错误，它们也可能导致重大的生产力损失。实际上，所有产生非常大结果的模拟都需要在保存数据之前以某种方式减少数据量——一种技术被称为有损压缩。本项目力求验证与有损压缩相结合的仿真的最终结果。这种方法对气候、宇宙学、流体动力学、天气和天体物理学等不同领域的科学模拟很有用——这些都是该项目的驱动因素。这个协作项目应用了外部算法观察者（EAO）的原则，其中科学应用程序的产品与复杂性低得多的替代函数的产品进行比较。使用三模冗余的变体来纠正损坏：如果检测到损坏，则执行第二个代理函数，并从最一致的两个结果中选择正确的值。这种新的在线检测/校正方法涉及对科学应用程序和替代函数的有损压缩结果进行近似比较。该项目探讨了代理函数、有损压缩器和近似比较技术的检测性能。该项目还探讨了如何选择代理、有损压缩和近似函数来优化用户设置的目标和约束。评估考虑了跨越不同计算方法的五种应用程序，产生具有I/O瓶颈的大型数据集，并涵盖了与NSF相关的各种科学问题领域。除了满足在上述领域工作的科学家的需求外，该项目还将提高本科生的研究经验。计划在2016年夏季举办一所专注于弹性的暑期学校，腐败检测/纠正将是一个主要主题。该项目还在主要科学会议上组织教程，包括在线检测/校正数值模拟。

Trusting scientific applications requires guaranteeing the validity of computed results. Unfortunately, many examples of scientific computations have led to incorrect results, sometimes with catastrophic consequences. Currently known validation techniques cover only a fraction of the possible corruptions that numerical simulation and data analytics applications may suffer during execution. As science processes grow in size and complexity, the reliability and validity of their constituent steps is increasingly difficult to ascertain. Assessing validity in the presence of potential data corruptions is a serious and insufficiently recognized problem. Corruption may occur at all levels of computing, from the hardware to the application. An important aspect of these corruptions is that until they are discovered, all executions are at risk of being corrupted silently. In some documented cases, months have elapsed between the discovery of a corruption and notification to users. In the meantime, a potentially large number of executions may be corrupted, and incorrect conclusions may result. It may be difficult, after the fact, to check whether executions have actually been corrupted or not, so that even if corruptions do not lead to mistakes, they may lead to significant productivity losses. Virtually all simulations producing very large results need to reduce their data volume in some way before saving it --one technique is called lossy compression. This project strives to validate the end result of the simulation coupled with lossy compression. This approach is useful for scientific simulations in such diverse areas as climate, cosmology, fluid dynamics, weather, and astrophysics --the drivers of this project. This collaborative project applies the principle of an external algorithmic observer (EAO), where the product of a scientific application is compared with that of a surrogate function of much lower complexity. Corruptions are corrected using a variation of triple modular redundancy: if a corruption is detected, a second surrogate function is executed, and the correct value is chosen from the two results that are most in agreement. This new online detection/correction approach involves approximate comparison of the lossy compressed results of the scientific application and the surrogate function. The project explores the detection performance of surrogate functions, lossy compressors, and approximate comparison techniques. The project also explores how to select the surrogate, lossy compression, and approximate functions to optimize objectives and constraints set by the users. The evaluation considers a set of five applications spanning different computational methods, producing large datasets with I/O bottlenecks, and covering a variety of science problem domains relevant to the NSF. In addition to serving the needs of scientists working in the fields listed above, this project will enhance the research experience of undergraduate students. A summer school focused on resilience is planned for summer 2016, and corruption detection/correction will be a major topic. The project is also organizing tutorials in major science conferences that include online detection/correction of numerical simulations.