Reliability for Fast, Persistent Memory Applications

这项研究是由两个技术趋势推动的，持久存储器和内存中计算。永久存储器技术的独特之处在于，它的访问方式类似于传统的DRAM，但它提供了像闪存一样的持久存储。包括英特尔和美光在内的几家制造商已经宣布，他们计划在2017年推出低成本、高性能、持久存储的产品。在内存计算中，大型数据集几乎全部保存在DRAM中，以便进行快速处理，这是由相对便宜的DRAM的可用性推动的。它支持构建快速应用程序，如键值存储和内存数据库。如今，这些内存中的应用程序构成了现代Web基础设施的核心，使网站能够以低延迟和高吞吐量为请求提供服务。 持久性内存非常适合内存中的应用程序。永久内存可以用来补充DRAM，它有助于确保数据的持久性。它相对较好的性能、低成本和高容量将使这些应用程序能够扩展到更大的数据集，并在更大的机器集群上运行。 在这个提案中，我的研究目标是解决这样一个问题：如何确保持久存储应用程序的可靠性？我有兴趣探索这些应用程序带来的所有可靠性挑战，包括确保正确性、可用性和灾难恢复。虽然最近的许多工作都集中在这些应用程序的性能上，但今天对这些可靠性问题的了解较少。 我计划探索几种提高持久性内存应用程序可靠性的方法，包括检测这些应用程序中的错误，为编写和测试这些应用程序提供语言支持，以及构建可靠和可伸缩的分布式应用程序。我预计，将这些方法应用于持久存储应用将带来新的研究结果，并产生重大影响。 如今，人们期望计算机系统能够可靠地工作并随时可用。然而，硬件故障和软件错误很常见，占系统停机时间的很大一部分。这些失败造成的经济影响估计高达数千亿美元。我希望这项研究的结果将有助于减少错误和故障的影响，从而支持可靠和快速的内存应用程序。

This research is driven by two technology trends, persistent memory and in-memory computing. Persistent memory technology is unique in that it is accessed similar to traditional DRAM, but it provides durable storage like Flash. Several manufactures, including Intel and Micron, have announced that they plan to make low cost, high-performance, persistent memory products available in 2017. In-memory computing, in which large data sets are kept almost entirely in DRAM for fast processing, is being driven by the availability of relatively inexpensive DRAM. It enables building fast applications such as key-value stores and in-memory databases. Today, these in-memory applications form the core of the modern web infrastructure, enabling web sites to serve requests with low latency and high throughput. Persistent memory is a natural fit for in-memory applications. Persistent memory can be used to supplement DRAM, and it helps ensure durability of data. Its relatively good performance, low cost and high capacity will enable scaling these applications to much larger data sets, and running them on larger clusters of machines. In this proposal, my research objective is to address the question: how can we ensure the reliability of persistent memory applications? I am interested in exploring all the reliability challenges raised by these applications, including ensuring correctness, availability and disaster recovery. While much recent work has focused on the performance of these applications, these reliability issues are less well understood today. I plan to explore several methods for improving the reliability of persistent memory applications, including detecting bugs in these applications, providing language support for writing and testing these applications, and building reliable and scalable distributed applications. I expect that applying these methods to persistent memory applications will lead to novel research results, with significant impact. Today, computer systems are expected to work reliably and be available at all times. Yet hardware failures and software bugs are common, accounting for a significant portion of system downtime. The economic impact of these failures is estimated to be in the hundreds of billions of dollars. I expect that the results of this research will help reduce the impact of bugs and failures, thereby enabling reliable and fast in-memory applications.